visit quad function example

Extend time allocation for testing on ruby from 45 to 60 minutes

.gitignore

@@ -211,7 +211,7 @@ miniapps/electromagnetics/joule
 miniapps/electromagnetics/Volta-AMR*
 miniapps/electromagnetics/Tesla-AMR*
 miniapps/electromagnetics/Maxwell-Parallel*
-miniapps/electromagnetics/Joule_*
+miniapps/electromagnetics/Joule_[0-9]*
 
 miniapps/gslib/field-diff
 miniapps/gslib/field-interp
@@ -267,9 +267,9 @@ miniapps/meshing/bounding-box*
 miniapps/meshing/jacobian-determinant*
 
 miniapps/mtop/parheat
-miniapps/mtop/ParHeat*
+miniapps/mtop/ParHeat/*
 miniapps/mtop/seqheat
-miniapps/mtop/SeqHeat*
+miniapps/mtop/SeqHeat/*
 
 miniapps/autodiff/paradiff
 miniapps/autodiff/seqadiff
@@ -277,7 +277,7 @@ miniapps/autodiff/seqtest
 miniapps/autodiff/par_example
 miniapps/autodiff/seq_example
 miniapps/autodiff/seq_test
-miniapps/autodiff/Exampl*
+miniapps/autodiff/Example/*
 
 miniapps/navier/navier_mms
 miniapps/navier/navier_kovasznay

.gitlab/configs/ruby-config.yml

@@ -52,4 +52,4 @@ variables:
     - echo ${JOBID}
     - echo ${MFEM_DATA_DIR}
     - echo ${SPEC}
-    - srun $( [[ -n "${JOBID}" ]] && echo "--jobid=${JOBID}" ) --reservation=ci -t 45 -N 1 tests/gitlab/build_and_test --spec "${SPEC}" --data-dir "${MFEM_DATA_DIR}" --data
+    - srun $( [[ -n "${JOBID}" ]] && echo "--jobid=${JOBID}" ) --reservation=ci -t 60 -N 1 tests/gitlab/build_and_test --spec "${SPEC}" --data-dir "${MFEM_DATA_DIR}" --data

CHANGELOG

@@ -46,6 +46,11 @@ GPU computing
   where host execution is always needed (e.g. when the DOFs array is small).
 - Introduced MFEM_FOREACH_THREAD_DIRECT, which directly maps loop tasks to GPU
   threads, assigning one task per thread.
+- Implemented a GPU-accelerated matrix-free AMR derefinement `GridFunction`
+  update operator. This supports mixed geometry meshes and variable order
+  spaces, and is the default derefinement operator constructed by
+  `FiniteElementSpace::Update` and `ParFiniteElementSpace::Update`.
+  The operator requires `FiniteElementSpace::Nonconforming() == true`.
 
 New and updated examples and miniapps
 -------------------------------------

CMakeLists.txt

@@ -598,14 +598,20 @@ set(MFEM_TPLS OPENMP HYPRE LAPACK BLAS SuperLUDist STRUMPACK METIS SuiteSparse
     BENCHMARK PARELAG TRIBOL MPI_CXX HIP HIPBLAS HIPSPARSE MOONOLITH BLITZ
     ALGOIM ENZYME)
 
-# Add all *_FOUND libraries in the variable TPL_LIBRARIES.
+# Add all created targets and *_FOUND libraries in the variables TPL_TARGETS and
+# TPL_LIBRARIES, respectively.
+set(TPL_TARGETS)
 set(TPL_LIBRARIES "")
 set(TPL_INCLUDE_DIRS "")
 foreach(TPL IN LISTS MFEM_TPLS)
-  if (${TPL}_FOUND)
+  if (${TPL}_FOUND OR TARGET ${TPL})
     message(STATUS "MFEM: using package ${TPL}")
-    list(APPEND TPL_LIBRARIES ${${TPL}_LIBRARIES})
-    list(APPEND TPL_INCLUDE_DIRS ${${TPL}_INCLUDE_DIRS})
+    if (TARGET ${TPL})
+      list(APPEND TPL_TARGETS ${TPL})
+    else()
+      list(APPEND TPL_LIBRARIES ${${TPL}_LIBRARIES})
+      list(APPEND TPL_INCLUDE_DIRS ${${TPL}_INCLUDE_DIRS})
+    endif()
   endif()
 endforeach(TPL)
 list(REVERSE TPL_LIBRARIES)
@@ -680,7 +686,10 @@ set(MFEM_INSTALL_DIR ${CMAKE_INSTALL_PREFIX})
 # Declaring the library
 mfem_add_library(mfem ${SOURCES} ${HEADERS} ${MASTER_HEADERS})
 # message(STATUS "TPL_LIBRARIES = ${TPL_LIBRARIES}")
-target_link_libraries(mfem PUBLIC ${TPL_LIBRARIES})
+target_link_libraries(mfem PUBLIC ${TPL_LIBRARIES} ${TPL_TARGETS})
+if (TPL_TARGETS)
+  add_dependencies(mfem ${TPL_TARGETS})
+endif()
 if (MINGW)
   target_link_libraries(mfem PRIVATE ws2_32)
 endif()

INSTALL

@@ -121,6 +121,11 @@ Parallel build:
    make -j 4
 (For METIS 5, see https://mfem.org/building/#parallel-build-using-metis-5)
 
+Parallel build with fetching of hypre and METIS:
+   mkdir <mfem-buil-dir> ; cd <mfem-build-dir>
+   cmake <mfem-source-dir> -DMFEM_USE_MPI=YES -DFETCH_TPLS=YES
+   make -j 4
+
 CUDA build:
    (this build requires CMake 3.17 or newer)
    mkdir <mfem-build-dir> ; cd <mfem-build-dir>
@@ -1074,6 +1079,9 @@ The following options are CMake specific:
 MFEM_ENABLE_TESTING  - Enable the ctest framework for testing.
 MFEM_ENABLE_EXAMPLES - Build all of the examples by default.
 MFEM_ENABLE_MINIAPPS - Build all of the miniapps by default.
+FETCH_TPLS           - Enable fetching of all supported third-party libraries.
+HYPRE_FETCH          - Enable fetching of hypre.
+METIS_FETCH          - Enable fetching of metis.
 
 External libraries (CMake):
 ---------------------------
@@ -1135,6 +1143,12 @@ The following built-in CMake packages are also used:
       set the <LIBNAME>_LIBRARIES option directly; the configuration option
       <LIBNAME>_DIR is not supported.
 
+The MFEM CMake build system also provides fetching (automated building) for the
+packages/libraries listed below. Note that when fetching is enabled, any related
+auto-detection functionality is disabled.
+
+ - HYPRE
+ - METIS
 
 Building without GNU make or CMake
 ==================================

config/cmake/modules/FindHYPRE.cmake

@@ -9,15 +9,18 @@
 # terms of the BSD-3 license. We welcome feedback and contributions, see file
 # CONTRIBUTING.md for details.
 
-# Defines the following variables:
+# Defines the following variables if fetching of TPLs is disabled (default):
 #   - HYPRE_FOUND
 #   - HYPRE_LIBRARIES
 #   - HYPRE_INCLUDE_DIRS
 #   - HYPRE_VERSION
 #   - HYPRE_USING_CUDA (internal)
 #   - HYPRE_USING_HIP (internal)
+# otherwise, the following are defined:
+#   - HYPRE (imported library target)
+#   - HYPRE_VERSION (cache variable)
 
-if (HYPRE_FOUND)
+if (HYPRE_FOUND OR TARGET HYPRE)
   if (HYPRE_USING_CUDA)
     find_package(CUDAToolkit REQUIRED)
   endif()
@@ -33,6 +36,55 @@ if (HYPRE_FOUND)
   endif()
 endif()
 
+if (HYPRE_FETCH OR FETCH_TPLS)
+  set(HYPRE_FETCH_VERSION 2.33.0)
+  add_library(HYPRE STATIC IMPORTED)
+  # set options and associated dependencies
+  set(CMAKE_OPTIONS)
+  list(APPEND CMAKE_OPTIONS -DCMAKE_BUILD_TYPE:STRING=${CMAKE_BUILD_TYPE})
+  if (MFEM_USE_CUDA)
+    list(APPEND CMAKE_OPTIONS -DHYPRE_WITH_CUDA:BOOL=ON)
+    find_package(CUDAToolkit REQUIRED)
+    target_link_libraries(HYPRE INTERFACE CUDA::cusparse CUDA::curand CUDA::cublas)
+  elseif (MFEM_USE_HIP)
+    list(APPEND CMAKE_OPTIONS -DHYPRE_WITH_HIP:BOOL=ON)
+    find_package(rocsparse REQUIRED)
+    find_package(rocrand REQUIRED)
+    target_link_libraries(HYPRE INTERFACE rocsparse rocrand)
+  endif()
+  if (MFEM_USE_SINGLE)
+    list(APPEND CMAKE_OPTIONS -DHYPRE_ENABLE_SINGLE:BOOL=ON)
+  endif()
+  # define external project and create future include directory so it is present
+  # to pass CMake checks at end of MFEM configuration step
+  message(STATUS "Will fetch HYPRE ${HYPRE_FETCH_VERSION} to be built with ${CMAKE_OPTIONS}")
+  set(PREFIX ${CMAKE_BINARY_DIR}/fetch/hypre)
+  include(ExternalProject)
+  ExternalProject_Add(hypre
+    GIT_REPOSITORY https://github.com/hypre-space/hypre.git
+    GIT_TAG v${HYPRE_FETCH_VERSION}
+    GIT_SHALLOW TRUE
+    UPDATE_DISCONNECTED TRUE
+    SOURCE_SUBDIR src
+    PREFIX ${PREFIX}
+    CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${PREFIX} -DCMAKE_INSTALL_LIBDIR:PATH=lib ${CMAKE_OPTIONS})
+  file(MAKE_DIRECTORY ${PREFIX}/include)
+  # set imported library target properties
+  add_dependencies(HYPRE hypre)
+  set_target_properties(HYPRE PROPERTIES
+    IMPORTED_LOCATION ${PREFIX}/lib/libHYPRE.a
+    INTERFACE_INCLUDE_DIRECTORIES ${PREFIX}/include)
+  # convert HYPRE version to integer
+  string(REGEX MATCHALL "[0-9]+" HYPRE_SPLIT_VERSION ${HYPRE_FETCH_VERSION})
+  list(GET HYPRE_SPLIT_VERSION 0 HYPRE_MAJOR_VERSION)
+  list(GET HYPRE_SPLIT_VERSION 1 HYPRE_MINOR_VERSION)
+  list(GET HYPRE_SPLIT_VERSION 2 HYPRE_PATCH_VERSION)
+  math(EXPR HYPRE_VERSION "10000*${HYPRE_MAJOR_VERSION} + 100*${HYPRE_MINOR_VERSION} + ${HYPRE_PATCH_VERSION}")
+  # set cache variables that would otherwise be set after mfem_find_package call
+  set(HYPRE_VERSION ${HYPRE_VERSION} CACHE STRING "HYPRE version." FORCE)
+  return()
+endif()
+
 include(MfemCmakeUtilities)
 mfem_find_package(HYPRE HYPRE HYPRE_DIR "include" "HYPRE.h" "lib" "HYPRE"
   "Paths to headers required by HYPRE." "Libraries required by HYPRE."

config/cmake/modules/FindMETIS.cmake

@@ -9,10 +9,38 @@
 # terms of the BSD-3 license. We welcome feedback and contributions, see file
 # CONTRIBUTING.md for details.
 
-# Defines the following variables:
+# Defines the following variables if fetching of TPLs is disabled (default):
 #   - METIS_FOUND
 #   - METIS_LIBRARIES
 #   - METIS_INCLUDE_DIRS
+#   - METIS_VERSION_5
+# otherwise, the following are defined:
+#   - METIS (imported library target)
+#   - METIS_VERSION_5 (cache variable)
+
+if (METIS_FETCH OR FETCH_TPLS)
+  set(METIS_FETCH_VERSION 4.0.3)
+  add_library(METIS STATIC IMPORTED)
+  # define external project
+  message(STATUS "Will fetch METIS ${METIS_FETCH_VERSION} to be built with default options")
+  set(PREFIX ${CMAKE_BINARY_DIR}/fetch/metis)
+  include(ExternalProject)
+  ExternalProject_Add(metis
+    GIT_REPOSITORY https://github.com/mfem/tpls
+    GIT_TAG b60352fbe9675d374b00828055e55be4584c7995 # tag from 1/16/25
+    GIT_SHALLOW TRUE
+    UPDATE_DISCONNECTED TRUE
+    PREFIX ${PREFIX}
+    CONFIGURE_COMMAND tar -xzf ../metis/metis-${METIS_FETCH_VERSION}-mac.tgz --strip=1
+    INSTALL_COMMAND mkdir -p ${PREFIX}/lib && cp libmetis.a ${PREFIX}/lib/)
+  # set imported library target properties
+  add_dependencies(METIS metis)
+  set_target_properties(METIS PROPERTIES
+    IMPORTED_LOCATION ${PREFIX}/lib/libmetis.a)
+  # set cache variables that would otherwise be set after mfem_find_package call
+  set(METIS_VERSION_5 FALSE CACHE BOOL "Is METIS version 5?")
+  return()
+endif()
 
 include(MfemCmakeUtilities)
 mfem_find_package(METIS METIS METIS_DIR "include;Lib" "metis.h"

config/defaults.cmake

@@ -89,6 +89,12 @@ option(MFEM_ENABLE_EXAMPLES "Build all of the examples" OFF)
 option(MFEM_ENABLE_MINIAPPS "Build all of the miniapps" OFF)
 option(MFEM_ENABLE_BENCHMARKS "Build all of the benchmarks" OFF)
 
+# Allow a user to specify fetching of certain third-party libraries instead of
+# searching for existing installations.
+option(FETCH_TPLS "Enable fetching of all supported third-party libraries" OFF)
+option(HYPRE_FETCH "Enable fetching of hypre" OFF)
+option(METIS_FETCH "Enable fetching of METIS" OFF)
+
 # Setting CXX/MPICXX on the command line or in user.cmake will overwrite the
 # autodetected C++ compiler.
 # set(CXX g++)

examples/visit_quadfunction.cpp

@@ -0,0 +1,231 @@
+#include "mfem.hpp"
+#include <fstream>
+#include <iostream>
+
+using namespace std;
+using namespace mfem;
+
+//setup intial condition
+int problem = 3;
+
+// Initial condition
+real_t my_function(const Vector &x);
+
+// Mesh bounding box
+Vector bb_min, bb_max;
+
+
+
+int main(int argc, char *argv[])
+{
+
+   Mpi::Init();
+   int num_procs = Mpi::WorldSize();
+   int myid = Mpi::WorldRank();
+
+   // 1. Parse command-line options.
+   const char *mesh_file = "../data/periodic-hexagon.mesh";
+   int order = 2;
+   bool pa = false;
+   bool fa = false;
+   const char *device_config = "cpu";
+   bool visualization = true;
+   bool algebraic_ceed = false;
+
+   OptionsParser args(argc, argv);
+   args.AddOption(&mesh_file, "-m", "--mesh",
+                  "Mesh file to use.");
+   args.AddOption(&order, "-o", "--order",
+                  "Finite element order (polynomial degree) or -1 for"
+                  " isoparametric space.");
+   args.AddOption(&pa, "-pa", "--partial-assembly", "-no-pa",
+                  "--no-partial-assembly", "Enable Partial Assembly.");
+   args.AddOption(&fa, "-fa", "--full-assembly", "-no-fa",
+                  "--no-full-assembly", "Enable Full Assembly.");
+   args.AddOption(&device_config, "-d", "--device",
+                  "Device configuration string, see Device::Configure().");
+   args.AddOption(&visualization, "-vis", "--visualization", "-no-vis",
+                  "--no-visualization",
+                  "Enable or disable GLVis visualization.");
+   args.Parse();
+
+   if (!args.Good())
+   {
+      if (Mpi::Root())
+      {
+         args.PrintUsage(cout);
+      }
+      return 1;
+   }
+   if (Mpi::Root())
+   {
+      args.PrintOptions(cout);
+   }
+
+
+   // 2. Enable hardware devices such as GPUs, and programming models such as
+   //    CUDA, OCCA, RAJA and OpenMP based on command line options.
+   Device device(device_config);
+   if (Mpi::Root()) { device.Print(); }
+
+   // 3. Read the mesh from the given mesh file. We can handle triangular,
+   //    quadrilateral, tetrahedral, hexahedral, surface and volume meshes with
+   //    the same code.
+   Mesh mesh(mesh_file, 1, 1);
+   int dim = mesh.Dimension();
+
+   // 4. Refine the mesh to increase the resolution. In this example we do
+   //    'ref_levels' of uniform refinement. We choose 'ref_levels' to be the
+   //    largest number that gives a final mesh with no more than 50,000
+   //    elements.
+   {
+     int ref_levels = 1; //(int)floor(log(50000./mesh.GetNE())/log(2.)/dim);
+
+      for (int l = 0; l < ref_levels; l++)
+      {
+         mesh.UniformRefinement();
+      }
+   }
+
+   ParMesh pmesh(MPI_COMM_WORLD, mesh);
+
+   pmesh.GetBoundingBox(bb_min, bb_max, max(order, 1));
+
+   const int elem_index = 0;
+   mfem::Element *el = pmesh.GetElement(elem_index);
+   mfem::Geometry::Type geom = el->GetGeometryType();
+
+   //Create finite element space
+   DG_FECollection fec(order, pmesh.Dimension());
+   ParFiniteElementSpace *pfespace = new ParFiniteElementSpace(&pmesh, &fec);
+
+
+   //specify integration rule
+   int intorder = (2*order-1);
+   //IntegrationRules irGL(geom, Quadrature1D::GaussLobatto);
+   const IntegrationRule *ir = &IntRules.Get(geom, 2*order);
+
+   //Create QuadratureSpace
+   QuadratureSpace qspace(pmesh, *ir);
+
+   QuadratureFunction quadfun(qspace);
+
+   std::cout<<"quadfun.Size() = "<<quadfun.Size()<<std::endl;
+   std::cout<<"pmesh.GetNE() * qpts "<<pmesh.GetNE() * ir->GetNPoints()<<std::endl;
+
+   //Loop over elements and quadrature points
+   int qf_idx = 0;
+   for (int el = 0; el < pmesh.GetNE(); el++)
+   {
+     ElementTransformation *trans = pmesh.GetElementTransformation(el);
+     for (int q = 0; q < ir->GetNPoints(); q++)
+     {
+       const IntegrationPoint &ip = ir->IntPoint(q);
+       Vector phys_point(pmesh.Dimension());
+       trans->Transform(ip, phys_point); // Map to physical coordinates
+
+       quadfun(qf_idx++) = my_function(phys_point);
+     }
+   }
+
+
+   {
+      int precision = 8;
+      cout.precision(precision);
+
+      ofstream omesh("main.mesh");
+      omesh.precision(precision);
+      pmesh.Print(omesh);
+      ofstream oquad("intial.qf");
+      oquad.precision(precision);
+      quadfun.Save(oquad);
+
+   }
+
+
+   //
+   //Create visit file
+   //
+   {
+     VisItDataCollection *pd = NULL;
+     pd = new VisItDataCollection("quad_fun", &pmesh);
+     pd->SetPrefixPath("Visit");
+     pd->RegisterQField("quad_field", &quadfun);
+     pd->SetLevelsOfDetail(order);
+     //pd->SetDataFormat(VTKFormat::BINARY);
+     //pd->SetHighOrderOutput(true);
+     pd->SetCycle(0);
+     pd->SetTime(0.0);
+     pd->Save();
+     delete pd;
+   }
+
+
+   // 14. Send the solution by socket to a GLVis server.
+   if (visualization && false)
+   {
+      char vishost[] = "localhost";
+      int  visport   = 19916;
+      socketstream sol_sock(vishost, visport);
+      sol_sock.precision(8);
+      sol_sock << "solution\n" << pmesh << quadfun << flush;
+   }
+
+   // 15. Free the used memory.
+
+
+   return 0;
+}
+
+// Initial condition
+real_t my_function(const Vector &x)
+{
+   int dim = x.Size();
+
+   // map to the reference [-1,1] domain
+   Vector X(dim);
+   for (int i = 0; i < dim; i++)
+   {
+      real_t center = (bb_min[i] + bb_max[i]) * 0.5;
+      X(i) = 2 * (x(i) - center) / (bb_max[i] - bb_min[i]);
+   }
+
+   switch (problem)
+   {
+      case 0:
+      case 1:
+      {
+         switch (dim)
+         {
+            case 1:
+               return exp(-40.*pow(X(0)-0.5,2));
+            case 2:
+            case 3:
+            {
+               real_t rx = 0.45, ry = 0.25, cx = 0., cy = -0.2, w = 10.;
+               if (dim == 3)
+               {
+                  const real_t s = (1. + 0.25*cos(2*M_PI*X(2)));
+                  rx *= s;
+                  ry *= s;
+               }
+               return ( std::erfc(w*(X(0)-cx-rx))*std::erfc(-w*(X(0)-cx+rx)) *
+                        std::erfc(w*(X(1)-cy-ry))*std::erfc(-w*(X(1)-cy+ry)) )/16;
+            }
+         }
+      }
+      case 2:
+      {
+         real_t x_ = X(0), y_ = X(1), rho, phi;
+         rho = std::hypot(x_, y_);
+         phi = atan2(y_, x_);
+         return pow(sin(M_PI*rho),2)*sin(3*phi);
+      }
+      case 3:
+      {
+         const real_t f = M_PI;
+         return 2 * sin(f*X(0))*sin(f*X(1));
+      }
+   }
+   return 0.0;
+}

fem/CMakeLists.txt

@@ -82,6 +82,8 @@ set(SRCS
   fe/fe_ser.cpp
   fe_coll.cpp
   fespace.cpp
+  derefmat_op.cpp
+  pderefmat_op.cpp
   geom.cpp
   gridfunc.cpp
   hybridization.cpp

fem/derefmat_op.cpp

@@ -0,0 +1,266 @@
+// 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 "derefmat_op.hpp"
+#include "fes_kernels.hpp"
+
+/// \cond DO_NOT_DOCUMENT
+namespace mfem
+{
+
+namespace internal
+{
+template <Ordering::Type Order, bool Atomic>
+static void DerefMultKernelImpl(const DerefineMatrixOp &op, const Vector &x,
+                                Vector &y)
+{
+   DerefineMatrixOpMultFunctor<Order, Atomic> func;
+   func.xptr = x.Read();
+   y.UseDevice();
+   y = 0.;
+   func.yptr = y.ReadWrite();
+   func.bsptr = op.block_storage.Read();
+   func.boptr = op.block_offsets.Read();
+   func.brptr = op.block_row_idcs_offsets.Read();
+   func.bcptr = op.block_col_idcs_offsets.Read();
+   func.rptr = op.row_idcs.Read();
+   func.cptr = op.col_idcs.Read();
+   func.vdims = op.fespace->GetVDim();
+   func.nblocks = op.block_offsets.Size();
+   func.width = op.Width() / func.vdims;
+   func.height = op.Height() / func.vdims;
+   func.Run(op.max_rows);
+}
+
+} // namespace internal
+
+DerefineMatrixOp::DerefineMatrixOp(FiniteElementSpace &fespace_, int old_ndofs,
+                                   const Table *old_elem_dof,
+                                   const Table *old_elem_fos)
+   : Operator(fespace_.GetVSize(), old_ndofs * fespace_.GetVDim()),
+     fespace(&fespace_)
+{
+   static Kernels kernels;
+   constexpr int max_team_size = 256;
+   /// TODO: Implement DofTransformation support
+
+   MFEM_VERIFY(fespace->Nonconforming(),
+               "Not implemented for conforming meshes.");
+   MFEM_VERIFY(old_ndofs, "Missing previous (finer) space.");
+   MFEM_VERIFY(fespace->GetNDofs() <= old_ndofs,
+               "Previous space is not finer.");
+
+   const CoarseFineTransformations &dtrans =
+      fespace->GetMesh()->ncmesh->GetDerefinementTransforms();
+
+   MFEM_ASSERT(dtrans.embeddings.Size() == old_elem_dof->Size(), "");
+
+   const bool is_dg = fespace->FEColl()->GetContType()
+                      == FiniteElementCollection::DISCONTINUOUS;
+   DenseMatrix localRVO; // for variable-order only
+
+   DenseTensor localR[Geometry::NumGeom];
+   int total_rows = 0;
+   int total_cols = 0;
+   block_offsets.SetSize(dtrans.embeddings.Size());
+   block_offsets.HostWrite();
+   if (fespace->IsVariableOrder())
+   {
+      // TODO: any potential for some compression here?
+      // determine storage size and offsets
+      block_offsets[0] = 0;
+      int total_size = 0;
+      for (int k = 0; k < dtrans.embeddings.Size(); ++k)
+      {
+         const Embedding &emb = dtrans.embeddings[k];
+         const FiniteElement *fe = fespace->GetFE(emb.parent);
+         const int ldof = fe->GetDof();
+         if (k + 1 < dtrans.embeddings.Size())
+         {
+            block_offsets[k + 1] = block_offsets[k] + ldof * ldof;
+         }
+         total_rows += ldof;
+         total_cols += ldof;
+         total_size += ldof * ldof;
+      }
+      block_storage.SetSize(total_size);
+   }
+   else
+   {
+      // compression scheme:
+      // block_offsets is the start of each block, potentially repeated
+      // only need to store localR for used shapes
+      Mesh::GeometryList elem_geoms(*fespace->GetMesh());
+
+      int geom_offsets[Geometry::NumGeom];
+      {
+         int size = 0;
+         for (int i = 0; i < elem_geoms.Size(); ++i)
+         {
+            fespace->GetLocalDerefinementMatrices(elem_geoms[i],
+                                                  localR[elem_geoms[i]]);
+            geom_offsets[elem_geoms[i]] = size;
+            size += localR[elem_geoms[i]].TotalSize();
+         }
+         block_storage.SetSize(size);
+         // copy blocks into block_storage
+         auto bs_ptr = block_storage.HostWrite();
+         for (int i = 0; i < elem_geoms.Size(); ++i)
+         {
+            std::copy(localR[elem_geoms[i]].Data(),
+                      localR[elem_geoms[i]].Data()
+                      + localR[elem_geoms[i]].TotalSize(),
+                      bs_ptr);
+            bs_ptr += localR[elem_geoms[i]].TotalSize();
+         }
+      }
+      for (int k = 0; k < dtrans.embeddings.Size(); ++k)
+      {
+         const Embedding &emb = dtrans.embeddings[k];
+         Geometry::Type geom =
+            fespace->GetMesh()->GetElementBaseGeometry(emb.parent);
+
+         auto size = localR[geom].SizeI() * localR[geom].SizeJ();
+         total_rows += localR[geom].SizeI();
+         total_cols += localR[geom].SizeJ();
+         // set block offsets and sizes
+         block_offsets[k] = geom_offsets[geom] + size * emb.matrix;
+      }
+   }
+   row_idcs.SetSize(total_rows);
+   row_idcs.HostWrite();
+   col_idcs.SetSize(total_cols);
+   col_idcs.HostWrite();
+   block_row_idcs_offsets.SetSize(dtrans.embeddings.Size() + 1);
+   block_row_idcs_offsets.HostWrite();
+   block_col_idcs_offsets.SetSize(dtrans.embeddings.Size() + 1);
+   block_col_idcs_offsets.HostWrite();
+   block_row_idcs_offsets[0] = 0;
+   block_col_idcs_offsets[0] = 0;
+
+   // compute index information
+   Array<int> dofs, old_dofs;
+   max_rows = 1;
+
+   {
+      Array<int> mark(fespace->GetNDofs());
+      mark = 0;
+      auto bs_ptr = block_storage.HostWrite();
+      int ridx = 0;
+      int cidx = 0;
+      int num_marked = 0;
+      for (int k = 0; k < dtrans.embeddings.Size(); k++)
+      {
+         const Embedding &emb = dtrans.embeddings[k];
+         Geometry::Type geom =
+            fespace->GetMesh()->GetElementBaseGeometry(emb.parent);
+
+         if (fespace->IsVariableOrder())
+         {
+            const FiniteElement *fe = fespace->GetFE(emb.parent);
+            const DenseTensor &pmats = dtrans.point_matrices[geom];
+            const int ldof = fe->GetDof();
+
+            IsoparametricTransformation isotr;
+            isotr.SetIdentityTransformation(geom);
+
+            localRVO.SetSize(ldof, ldof);
+            isotr.SetPointMat(pmats(emb.matrix));
+            // Local restriction is size ldofxldof assuming that the parent
+            // and child are of same polynomial order.
+            fe->GetLocalRestriction(isotr, localRVO);
+            // copy block
+            auto size = localRVO.Height() * localRVO.Width();
+            std::copy(localRVO.Data(), localRVO.Data() + size, bs_ptr);
+            bs_ptr += size;
+         }
+         DenseMatrix &lR =
+            fespace->IsVariableOrder() ? localRVO : localR[geom](emb.matrix);
+         block_row_idcs_offsets[k + 1] =
+            block_row_idcs_offsets[k] + lR.Height();
+         block_col_idcs_offsets[k + 1] = block_col_idcs_offsets[k] + lR.Width();
+         max_rows = std::max(lR.Height(), max_rows);
+         // index information
+         fespace->elem_dof->GetRow(emb.parent, dofs);
+         old_elem_dof->GetRow(k, old_dofs);
+         MFEM_VERIFY(old_dofs.Size() == dofs.Size(),
+                     "Parent and child must have same #dofs.");
+         for (int i = 0; i < lR.Height(); ++i, ++ridx)
+         {
+            if (!std::isfinite(lR(i, 0)))
+            {
+               row_idcs[ridx] = INT_MAX;
+               continue;
+            }
+            int r = dofs[i];
+            int m = (r >= 0) ? r : (-1 - r);
+            if (is_dg || !mark[m])
+            {
+               row_idcs[ridx] = r;
+               mark[m] = 1;
+               ++num_marked;
+            }
+            else
+            {
+               row_idcs[ridx] = INT_MAX;
+            }
+         }
+         for (int i = 0; i < lR.Width(); ++i, ++cidx)
+         {
+            col_idcs[cidx] = old_dofs[i];
+         }
+      }
+      if (!is_dg && !fespace->IsVariableOrder())
+      {
+         MFEM_VERIFY(num_marked * fespace->GetVDim() == Height(),
+                     "internal error: not all rows were set.");
+      }
+   }
+   // if not using GPU, set max_rows/max_cols to zero
+   if (Device::Allows(Backend::DEVICE_MASK))
+   {
+      max_rows = std::min(max_rows, max_team_size);
+   }
+   else
+   {
+      max_rows = 1;
+   }
+}
+
+void DerefineMatrixOp::Mult(const Vector &x, Vector &y) const
+{
+   const bool is_dg = fespace->FEColl()->GetContType()
+                      == FiniteElementCollection::DISCONTINUOUS;
+   // DG needs atomic summation
+   MultKernel::Run(fespace->GetOrdering(), is_dg, *this, x, y);
+}
+
+DerefineMatrixOp::Kernels::Kernels()
+{
+   MultKernel::Specialization<Ordering::byNODES, false>::Add();
+   MultKernel::Specialization<Ordering::byVDIM, false>::Add();
+   MultKernel::Specialization<Ordering::byNODES, true>::Add();
+   MultKernel::Specialization<Ordering::byVDIM, true>::Add();
+}
+
+template <Ordering::Type Order, bool Atomic>
+DerefineMatrixOp::MultKernelType DerefineMatrixOp::MultKernel::Kernel()
+{
+   return internal::DerefMultKernelImpl<Order, Atomic>;
+}
+
+DerefineMatrixOp::MultKernelType
+DerefineMatrixOp::MultKernel::Fallback(Ordering::Type, bool)
+{
+   MFEM_ABORT("invalid MultKernel parameters");
+}
+} // namespace mfem
+/// \endcond DO_NOT_DOCUMENT

fem/derefmat_op.hpp

@@ -0,0 +1,65 @@
+// 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.
+
+#ifndef MFEM_DEREFMAT_OP
+#define MFEM_DEREFMAT_OP
+
+#include "fespace.hpp"
+
+#include "kernel_dispatch.hpp"
+
+namespace mfem
+{
+
+/// \cond DO_NOT_DOCUMENT
+
+struct DerefineMatrixOp : public Operator
+{
+   FiniteElementSpace *fespace;
+   /// offsets into block_storage
+   Array<int> block_offsets;
+   /// offsets into row_idcs
+   Array<int> block_row_idcs_offsets;
+   /// offsets into col_idcs
+   Array<int> block_col_idcs_offsets;
+   /// mapping for row dofs, INT_MAX indicates the block row should be ignored.
+   /// negative means the row data should be negated.
+   Array<int> row_idcs;
+   /// mapping for col dofs, negative means the col data should be negated.
+   Array<int> col_idcs;
+   /// dense block matrices which can be reused to construct the full matrix
+   /// operation. These are stored contiguously and blocks have no restrictions
+   /// on shape (can be rectangle and differ from block to block).
+   Vector block_storage;
+   /// maximum height of any block in block_storage for GPU
+   /// parallelization, or 1 for CPU runs.
+   int max_rows;
+
+   using MultKernelType = void (*)(const DerefineMatrixOp &, const Vector &,
+                                   Vector &);
+   /// template args: ordering, atomic
+   MFEM_REGISTER_KERNELS(MultKernel, MultKernelType, (Ordering::Type, bool));
+
+   struct Kernels
+   {
+      Kernels();
+   };
+
+   void Mult(const Vector &x, Vector &y) const;
+
+   DerefineMatrixOp(FiniteElementSpace &fespace_, int old_ndofs,
+                    const Table *old_elem_dof, const Table *old_elem_fos);
+};
+
+/// \endcond DO_NOT_DOCUMENT
+
+} // namespace mfem
+#endif

fem/fes_kernels.hpp

@@ -0,0 +1,249 @@
+// 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.
+
+#ifndef MFEM_FES_KERNELS_HPP
+#define MFEM_FES_KERNELS_HPP
+
+#include "../general/forall.hpp"
+
+#include <climits>
+
+namespace mfem
+{
+/// \cond DO_NOT_DOCUMENT
+namespace internal
+{
+
+///
+/// Implements matrix-vector multiply $y = A x$ for a sparse matrix composed of
+/// a sum of smaller dense blocks. There is additional permutation/sign
+/// information associated with each block. The base class only implements
+/// helper routines such as computing block widths, index into x, index into y,
+/// and column in A given sub-block information.
+/// @sa DerefineMatrixOpMultFunctor
+///
+/// @tparam Order vdim ordering for x and y. Note that for Diag = false this is
+/// ignored for x as x has a special interleaved order.
+/// @tparam Base used for the curious recurring template pattern (CRTP) so the
+/// base class can access child class fields without virtual functions
+/// @tparam Diag true if this corresponds to the diagonal block (coarse element
+/// and fine element are on our rank), false otherwise (coarse element is on our
+/// rank, fine element is on a different rank).
+///
+template <Ordering::Type Order, class Base, bool Diag = true>
+struct DerefineMatrixOpFunctorBase;
+
+template <class Base>
+struct DerefineMatrixOpFunctorBase<Ordering::byNODES, Base, true>
+{
+   /// block column indices offsets
+   const int *bcptr;
+   /// column indices
+   const int *cptr;
+
+   int MFEM_HOST_DEVICE BlockWidth(int k) const
+   {
+      return bcptr[k + 1] - bcptr[k];
+   }
+
+   void MFEM_HOST_DEVICE Col(int j, int k, int &col, int &sign) const
+   {
+      col = cptr[bcptr[k] + j];
+      if (col < 0)
+      {
+         col = -1 - col;
+         sign = -sign;
+      }
+   }
+
+   int MFEM_HOST_DEVICE IndexX(int col, int vdim, int) const
+   {
+      return col + vdim * static_cast<const Base *>(this)->width;
+   }
+   int MFEM_HOST_DEVICE IndexY(int row, int vdim) const
+   {
+      return row + vdim * static_cast<const Base *>(this)->height;
+   }
+};
+
+template <class Base>
+struct DerefineMatrixOpFunctorBase<Ordering::byVDIM, Base, true>
+{
+   /// block column indices offsets
+   const int *bcptr;
+   /// column indices
+   const int *cptr;
+
+   int MFEM_HOST_DEVICE BlockWidth(int k) const
+   {
+      return bcptr[k + 1] - bcptr[k];
+   }
+
+   void MFEM_HOST_DEVICE Col(int j, int k, int &col, int &sign) const
+   {
+      col = cptr[bcptr[k] + j];
+      if (col < 0)
+      {
+         col = -1 - col;
+         sign = -sign;
+      }
+   }
+
+   int MFEM_HOST_DEVICE IndexX(int col, int vdim, int) const
+   {
+      return vdim + col * static_cast<const Base *>(this)->vdims;
+   }
+   int MFEM_HOST_DEVICE IndexY(int row, int vdim) const
+   {
+      return vdim + row * static_cast<const Base *>(this)->vdims;
+   }
+};
+
+template <class Base>
+struct DerefineMatrixOpFunctorBase<Ordering::byNODES, Base, false>
+{
+   /// receive segment offsets
+   const int *segptr;
+   /// receive segment index
+   const int *rsptr;
+   /// off-diagonal block column offsets
+   const int *coptr;
+   /// off-diagonal block widths
+   const int *bwptr;
+
+   int MFEM_HOST_DEVICE BlockWidth(int k) const { return bwptr[k]; }
+
+   void MFEM_HOST_DEVICE Col(int j, int k, int &col, int &sign) const
+   {
+      col = coptr[k] + j;
+   }
+
+   int MFEM_HOST_DEVICE IndexX(int col, int vdim, int k) const
+   {
+      int tmp = rsptr[k];
+      int segwidth = segptr[tmp + 1] - segptr[tmp];
+      return segptr[tmp] * static_cast<const Base *>(this)->vdims + col +
+             vdim * segwidth;
+   }
+   int MFEM_HOST_DEVICE IndexY(int row, int vdim) const
+   {
+      return row + vdim * static_cast<const Base *>(this)->height;
+   }
+};
+
+template <class Base>
+struct DerefineMatrixOpFunctorBase<Ordering::byVDIM, Base, false>
+{
+   /// receive segment offsets
+   const int *segptr;
+   /// receive segment index
+   const int *rsptr;
+   /// off-diagonal block column offsets
+   const int *coptr;
+   /// off-diagonal block widths
+   const int *bwptr;
+
+   int MFEM_HOST_DEVICE BlockWidth(int k) const { return bwptr[k]; }
+
+   void MFEM_HOST_DEVICE Col(int j, int k, int &col, int &sign) const
+   {
+      col = coptr[k] + j;
+   }
+
+   int MFEM_HOST_DEVICE IndexX(int col, int vdim, int k) const
+   {
+      int tmp = rsptr[k];
+      int segwidth = segptr[tmp + 1] - segptr[tmp];
+      return segptr[tmp] * static_cast<const Base *>(this)->vdims + col +
+             vdim * segwidth;
+   }
+   int MFEM_HOST_DEVICE IndexY(int row, int vdim) const
+   {
+      return vdim + row * static_cast<const Base *>(this)->vdims;
+   }
+};
+
+/// internally used to implement the derefinement operator Mult diagonal
+/// block
+template <Ordering::Type Order, bool Atomic, bool Diag = true>
+struct DerefineMatrixOpMultFunctor
+   : public DerefineMatrixOpFunctorBase<
+     Order, DerefineMatrixOpMultFunctor<Order, Atomic, Diag>, Diag>
+{
+   const real_t *xptr;
+   real_t *yptr;
+   /// block storage
+   const real_t *bsptr;
+   /// block offsets
+   const int *boptr;
+   /// block row index offsets
+   const int *brptr;
+   /// row indices
+   const int *rptr;
+
+   // number of blocks
+   int nblocks;
+   // number of components
+   int vdims;
+   /// overall operator height (for vdim = 1)
+   int height;
+   /// overall operator width (for vdim = 1)
+   int width;
+   void MFEM_HOST_DEVICE operator()(int kidx) const
+   {
+      int k = kidx % nblocks;
+      int vdim = kidx / nblocks;
+
+      int block_height = brptr[k + 1] - brptr[k];
+      int block_width = this->BlockWidth(k);
+      MFEM_FOREACH_THREAD(i, x, block_height)
+      {
+         int row = rptr[brptr[k] + i];
+         int rsign = 1;
+         if (row < 0)
+         {
+            row = -1 - row;
+            rsign = -1;
+         }
+         if (row < INT_MAX)
+         {
+            // row not marked as unused
+            real_t sum = 0;
+            for (int j = 0; j < block_width; ++j)
+            {
+               int col, sign = rsign;
+               this->Col(j, k, col, sign);
+               sum += sign * bsptr[boptr[k] + i + j * block_height] *
+                      xptr[this->IndexX(col, vdim, k)];
+            }
+#if defined(__CUDA_ARCH__) or defined(__HIP_DEVICE_COMPILE__)
+            if (Atomic)
+            {
+               atomicAdd(yptr + this->IndexY(row, vdim), sum);
+            }
+            else
+#endif
+            {
+               yptr[this->IndexY(row, vdim)] += sum;
+            }
+         }
+      }
+   }
+
+   /// N is the max block row size (doesn't have to be a power of 2)
+   void Run(int N) const { forall_2D(nblocks * vdims, N, 1, *this); }
+};
+
+} // namespace internal
+/// \endcond DO_NOT_DOCUMENT
+} // namespace mfem
+
+#endif

fem/fespace.cpp

@@ -17,6 +17,9 @@
 #include "fem.hpp"
 #include "ceed/interface/util.hpp"
 
+#include "derefmat_op.hpp"
+
+#include <algorithm>
 #include <cmath>
 #include <cstdarg>
 
@@ -24,9 +27,9 @@ using namespace std;
 
 namespace mfem
 {
-
-template <> void Ordering::
-DofsToVDofs<Ordering::byNODES>(int ndofs, int vdim, Array<int> &dofs)
+template <>
+void Ordering::DofsToVDofs<Ordering::byNODES>(int ndofs, int vdim,
+                                              Array<int> &dofs)
 {
    // static method
    int size = dofs.Size();
@@ -40,8 +43,9 @@ DofsToVDofs<Ordering::byNODES>(int ndofs, int vdim, Array<int> &dofs)
    }
 }
 
-template <> void Ordering::
-DofsToVDofs<Ordering::byVDIM>(int ndofs, int vdim, Array<int> &dofs)
+template <>
+void Ordering::DofsToVDofs<Ordering::byVDIM>(int ndofs, int vdim,
+                                             Array<int> &dofs)
 {
    // static method
    int size = dofs.Size();
@@ -55,7 +59,6 @@ DofsToVDofs<Ordering::byVDIM>(int ndofs, int vdim, Array<int> &dofs)
    }
 }
 
-
 FiniteElementSpace::FiniteElementSpace()
    : mesh(NULL), fec(NULL), vdim(0), ordering(Ordering::byNODES),
      ndofs(0), nvdofs(0), nedofs(0), nfdofs(0), nbdofs(0),
@@ -4244,7 +4247,11 @@ void FiniteElementSpace::Update(bool want_transform)
          case Mesh::DEREFINE:
          {
             BuildConformingInterpolation();
+#if 0
             Th.Reset(DerefinementMatrix(old_ndofs, old_elem_dof, old_elem_fos));
+#else
+            Th.Reset(new DerefineMatrixOp(*this, old_ndofs, old_elem_dof, old_elem_fos));
+#endif
             if (IsVariableOrder())
             {
                if (cP && cR_hp)

fem/fespace.hpp

@@ -113,7 +113,7 @@ class QuadratureSpace;
 class QuadratureInterpolator;
 class FaceQuadratureInterpolator;
 class PRefinementTransferOperator;
-
+struct DerefineMatrixOp;
 
 /** @brief Class FiniteElementSpace - responsible for providing FEM view of the
     mesh, mainly managing the set of degrees of freedom.
@@ -246,6 +246,7 @@ class FiniteElementSpace
    friend class PRefinementTransferOperator;
    friend void Mesh::Swap(Mesh &, bool);
    friend class LORBase;
+   friend struct DerefineMatrixOp;
 
 protected:
    /// The mesh that FE space lives on (not owned).

fem/pderefmat_op.cpp

@@ -0,0 +1,591 @@
+// 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 "pderefmat_op.hpp"
+
+#ifdef MFEM_USE_MPI
+
+#include "fes_kernels.hpp"
+/// \cond DO_NOT_DOCUMENT
+namespace mfem
+{
+namespace internal
+{
+template <Ordering::Type Order, bool Atomic>
+static void ParDerefMultKernelImpl(const ParDerefineMatrixOp &op,
+                                   const Vector &x, Vector &y)
+{
+   // pack sends
+   if (op.xghost_send.Size())
+   {
+      auto src = x.Read();
+      auto idcs = op.send_permutations.Read();
+      auto dst = Device::GetGPUAwareMPI() ? op.xghost_send.Write()
+                 : op.xghost_send.HostWrite();
+      auto vdims = op.fespace->GetVDim();
+      auto sptr = op.send_segment_idcs.Read();
+      auto lptr = op.send_segments.Read();
+      auto old_ndofs = x.Size() / vdims;
+
+      forall(op.send_permutations.Size(), [=] MFEM_HOST_DEVICE(int i)
+      {
+         int seg = sptr[i];
+         int width = lptr[seg + 1] - lptr[seg];
+         auto tdst = dst + i + lptr[seg] * vdims;
+         int sign = 1;
+         int col = idcs[i];
+         if (col < 0)
+         {
+            sign = -1;
+            col = -1 - col;
+         }
+         for (int vdim = 0; vdim < vdims; ++vdim)
+         {
+            tdst[vdim * width] =
+               sign
+               * src[Order == Ordering::byNODES ? (col + vdim * old_ndofs)
+                           : (col * vdims + vdim)];
+         }
+      });
+      // TODO: is this needed so we can send the packed data correctly?
+      // unclear for GPU-aware MPI, definitely required otherwise
+      MFEM_DEVICE_SYNC;
+   }
+   // initialize off-diagonal receive and send
+   op.requests.clear();
+   if (op.xghost_recv.Size())
+   {
+      auto vdims = op.fespace->GetVDim();
+      auto rcv = Device::GetGPUAwareMPI() ? op.xghost_recv.Write()
+                 : op.xghost_recv.HostWrite();
+      for (int i = 0; i < op.recv_ranks.Size(); ++i)
+      {
+         op.requests.emplace_back();
+         MPI_Irecv(rcv + op.recv_segments[i] * vdims,
+                   (op.recv_segments[i + 1] - op.recv_segments[i]) * vdims,
+                   MPITypeMap<real_t>::mpi_type, op.recv_ranks[i],
+                   MessageTag::DEREFINEMENT_MATRIX_CONSTRUCTION_DATA,
+                   op.fespace->GetComm(), &op.requests.back());
+      }
+   }
+   if (op.xghost_send.Size())
+   {
+      auto vdims = op.fespace->GetVDim();
+      // only is a GPU mem ptr if GPU-aware MPI is enabled
+      auto dst = Device::GetGPUAwareMPI() ? op.xghost_send.Write()
+                 : op.xghost_send.HostWrite();
+      for (int i = 0; i < op.send_ranks.Size(); ++i)
+      {
+         op.requests.emplace_back();
+         MPI_Isend(dst + op.send_segments[i] * vdims,
+                   (op.send_segments[i + 1] - op.send_segments[i]) * vdims,
+                   MPITypeMap<real_t>::mpi_type, op.send_ranks[i],
+                   MessageTag::DEREFINEMENT_MATRIX_CONSTRUCTION_DATA,
+                   op.fespace->GetComm(), &op.requests.back());
+      }
+   }
+   {
+      // diagonal
+      DerefineMatrixOpMultFunctor<Order, Atomic, true> func;
+      func.xptr = x.Read();
+      y.UseDevice();
+      y = 0.;
+      func.yptr = y.ReadWrite();
+      func.bsptr = op.block_storage.Read();
+      func.boptr = op.block_offsets.Read();
+      func.brptr = op.block_row_idcs_offsets.Read();
+      func.bcptr = op.block_col_idcs_offsets.Read();
+      func.rptr = op.row_idcs.Read();
+      func.cptr = op.col_idcs.Read();
+      func.vdims = op.fespace->GetVDim();
+      func.nblocks = op.block_offsets.Size();
+      func.width = op.Width() / func.vdims;
+      func.height = op.Height() / func.vdims;
+      func.Run(op.max_rows);
+   }
+   // wait for comm to finish, if any
+   if (op.requests.size())
+   {
+      MPI_Waitall(op.requests.size(), op.requests.data(), MPI_STATUSES_IGNORE);
+      if (op.xghost_recv.Size())
+      {
+         // off-diagonal kernel
+         DerefineMatrixOpMultFunctor<Order, Atomic, false> func;
+         // directly read from host-pinned memory if not using GPU-aware MPI
+         func.xptr = Device::GetGPUAwareMPI() ? op.xghost_recv.Read()
+                     : op.xghost_recv.HostRead();
+         func.yptr = y.ReadWrite();
+         func.bsptr = op.block_storage.Read();
+         func.boptr = op.off_diag_block_offsets.Read();
+         func.brptr = op.block_off_diag_row_idcs_offsets.Read();
+         func.rsptr = op.recv_segment_idcs.Read();
+         func.segptr = op.recv_segments.Read();
+         func.coptr = op.block_off_diag_col_offsets.Read();
+         func.bwptr = op.block_off_diag_widths.Read();
+         func.rptr = op.row_off_diag_idcs.Read();
+         func.vdims = op.fespace->GetVDim();
+         func.nblocks = op.off_diag_block_offsets.Size();
+         func.width = op.xghost_recv.Size() / func.vdims;
+         func.height = op.Height() / func.vdims;
+         func.Run(op.max_rows);
+      }
+   }
+}
+} // namespace internal
+
+template <Ordering::Type Order, bool Atomic>
+ParDerefineMatrixOp::MultKernelType ParDerefineMatrixOp::MultKernel::Kernel()
+{
+   return internal::ParDerefMultKernelImpl<Order, Atomic>;
+}
+
+ParDerefineMatrixOp::MultKernelType
+ParDerefineMatrixOp::MultKernel::Fallback(Ordering::Type, bool)
+{
+   MFEM_ABORT("invalid MultKernel parameters");
+}
+
+ParDerefineMatrixOp::Kernels::Kernels()
+{
+   MultKernel::Specialization<Ordering::byNODES, false>::Add();
+   MultKernel::Specialization<Ordering::byVDIM, false>::Add();
+   MultKernel::Specialization<Ordering::byNODES, true>::Add();
+   MultKernel::Specialization<Ordering::byVDIM, true>::Add();
+}
+
+void ParDerefineMatrixOp::Mult(const Vector &x, Vector &y) const
+{
+   const bool is_dg = fespace->FEColl()->GetContType()
+                      == FiniteElementCollection::DISCONTINUOUS;
+   // DG needs atomic summation
+   MultKernel::Run(fespace->GetOrdering(), is_dg, *this, x, y);
+   // use this to prevent xghost* from being re-purposed for subsequent Mult
+   // calls
+   MFEM_DEVICE_SYNC;
+}
+
+ParDerefineMatrixOp::ParDerefineMatrixOp(ParFiniteElementSpace &fespace_,
+                                         int old_ndofs,
+                                         const Table *old_elem_dof,
+                                         const Table *old_elem_fos)
+   : Operator(fespace_.GetVSize(), old_ndofs * fespace_.GetVDim()),
+     fespace(&fespace_)
+{
+   static Kernels kernels;
+   constexpr int max_team_size = 256;
+
+   const int NRanks = fespace->GetNRanks();
+
+   const int nrk = HYPRE_AssumedPartitionCheck() ? 2 : NRanks;
+
+   MFEM_VERIFY(fespace->Nonconforming(),
+               "Not implemented for conforming meshes.");
+   MFEM_VERIFY(fespace->old_dof_offsets[nrk],
+               "Missing previous (finer) space.");
+
+   const int MyRank = fespace->GetMyRank();
+   ParNCMesh *old_pncmesh = fespace->GetParMesh()->pncmesh;
+   const CoarseFineTransformations &dtrans =
+      old_pncmesh->GetDerefinementTransforms();
+   const Array<int> &old_ranks = old_pncmesh->GetDerefineOldRanks();
+
+   const bool is_dg = fespace->FEColl()->GetContType()
+                      == FiniteElementCollection::DISCONTINUOUS;
+   DenseMatrix localRVO; // for variable-order only
+
+   DenseTensor localR[Geometry::NumGeom];
+   int diag_rows = 0;
+   int off_diag_rows = 0;
+   int diag_cols = 0;
+
+   auto get_ldofs = [&](int k) -> int
+   {
+      const Embedding &emb = dtrans.embeddings[k];
+      if (fespace->IsVariableOrder())
+      {
+         const FiniteElement *fe = fespace->GetFE(emb.parent);
+         return fe->GetDof();
+      }
+      else
+      {
+         Geometry::Type geom =
+         fespace->GetParMesh()->GetElementBaseGeometry(emb.parent);
+         return fespace->FEColl()->FiniteElementForGeometry(geom)->GetDof();
+      }
+   };
+   Array<int> dofs, old_dofs;
+   max_rows = 1;
+   // first pass:
+   // - determine memory block lengths
+   // - identify dofs in x we need to send/receive
+   // don't need to send the indices, fine rank will re-arrange and sign
+   // change x before transmitting the ghost data
+
+   // key: coarse rank to send to
+   // value: old dofs to send (with sign)
+   std::map<int, std::vector<int>> to_send;
+   // key: fine rank
+   // value: indices into dtrans.embeddings
+   std::map<int, std::vector<int>> od_ks;
+   // key: fine rank
+   // value: recv segment length
+   std::map<int, int> od_seg_lens;
+   int send_len = 0;
+   int recv_len = 0;
+   // size of block_storage, if fespace->IsVariableOrder()
+   // otherwise unused
+   int total_size = 0;
+   int num_diagonal_blocks = 0;
+   int num_offdiagonal_blocks = 0;
+   for (int k = 0; k < dtrans.embeddings.Size(); ++k)
+   {
+      const Embedding &emb = dtrans.embeddings[k];
+      int fine_rank = old_ranks[k];
+      int coarse_rank = (emb.parent < 0) ? (-1 - emb.parent)
+                        : old_pncmesh->ElementRank(emb.parent);
+      if (coarse_rank != MyRank && fine_rank == MyRank)
+      {
+         // this rank needs to send data in x to course_rank
+         old_elem_dof->GetRow(k, old_dofs);
+         auto &tmp = to_send[coarse_rank];
+         send_len += old_dofs.Size();
+         for (int i = 0; i < old_dofs.Size(); ++i)
+         {
+            tmp.emplace_back(old_dofs[i]);
+         }
+      }
+      else if (coarse_rank == MyRank && fine_rank != MyRank)
+      {
+         // this rank needs to receive data in x from fine_rank
+         MFEM_ASSERT(emb.parent >= 0, "");
+         auto ldofs = get_ldofs(k);
+         off_diag_rows += ldofs;
+         recv_len += ldofs;
+         od_ks[fine_rank].emplace_back(k);
+         od_seg_lens[fine_rank] += ldofs;
+         ++num_offdiagonal_blocks;
+         if (fespace->IsVariableOrder())
+         {
+            total_size += ldofs * ldofs;
+         }
+      }
+      else if (coarse_rank == MyRank && fine_rank == MyRank)
+      {
+         MFEM_ASSERT(emb.parent >= 0, "");
+         // diagonal
+         ++num_diagonal_blocks;
+         auto ldofs = get_ldofs(k);
+         diag_rows += ldofs;
+         diag_cols += ldofs;
+         if (fespace->IsVariableOrder())
+         {
+            total_size += ldofs * ldofs;
+         }
+      }
+   }
+   send_segments.SetSize(to_send.size() + 1);
+   send_segments.HostWrite();
+   send_ranks.SetSize(to_send.size());
+   send_ranks.HostWrite();
+   {
+      int idx = 0;
+      send_segments[0] = 0;
+      for (auto &tmp : to_send)
+      {
+         send_ranks[idx] = tmp.first;
+         send_segments[idx + 1] = send_segments[idx] + tmp.second.size();
+         ++idx;
+      }
+   }
+   recv_segment_idcs.SetSize(off_diag_rows);
+   recv_segment_idcs.HostWrite();
+   recv_segments.SetSize(od_ks.size() + 1);
+   recv_segments.HostWrite();
+   recv_ranks.SetSize(od_ks.size());
+   recv_ranks.HostWrite();
+
+   // set sizes
+   row_idcs.SetSize(diag_rows);
+   row_idcs.HostWrite();
+   row_off_diag_idcs.SetSize(off_diag_rows);
+   row_off_diag_idcs.HostWrite();
+   col_idcs.SetSize(diag_cols);
+   col_idcs.HostWrite();
+   block_row_idcs_offsets.SetSize(num_diagonal_blocks + 1);
+   block_row_idcs_offsets.HostWrite();
+   block_col_idcs_offsets.SetSize(num_diagonal_blocks + 1);
+   block_col_idcs_offsets.HostWrite();
+   block_off_diag_row_idcs_offsets.SetSize(num_offdiagonal_blocks + 1);
+   block_off_diag_row_idcs_offsets.HostWrite();
+   block_off_diag_col_offsets.SetSize(num_offdiagonal_blocks);
+   block_off_diag_col_offsets.HostWrite();
+   block_off_diag_widths.SetSize(num_offdiagonal_blocks);
+   block_off_diag_widths.HostWrite();
+   pack_col_idcs.SetSize(send_len);
+   // memory manager doesn't appear to have a graceful fallback for
+   // HOST_PINNED if not built with CUDA or HIP
+#if defined(MFEM_USE_CUDA) or defined(MFEM_USE_HIP)
+   xghost_send.SetSize(send_len * fespace->GetVDim(),
+                       Device::GetGPUAwareMPI() ? MemoryType::DEFAULT
+                       : MemoryType::HOST_PINNED);
+   xghost_recv.SetSize(recv_len * fespace->GetVDim(),
+                       Device::GetGPUAwareMPI() ? MemoryType::DEFAULT
+                       : MemoryType::HOST_PINNED);
+#else
+   xghost_send.SetSize(send_len * fespace->GetVDim());
+   xghost_recv.SetSize(recv_len * fespace->GetVDim());
+#endif
+   send_permutations.SetSize(send_len);
+   send_segment_idcs.SetSize(send_len);
+   block_offsets.SetSize(num_diagonal_blocks);
+   block_offsets.HostWrite();
+   off_diag_block_offsets.SetSize(num_offdiagonal_blocks);
+   off_diag_block_offsets.HostWrite();
+   int geom_offsets[Geometry::NumGeom];
+   real_t *bs_ptr;
+
+   if (fespace->IsVariableOrder())
+   {
+      block_storage.SetSize(total_size);
+      bs_ptr = block_storage.HostWrite();
+      // compute block data later
+   }
+   else
+   {
+      // compression scheme:
+      // block_offsets is the start of each block, potentially repeated
+      // only need to store localR for used shapes
+      Mesh::GeometryList elem_geoms(*fespace->GetMesh());
+
+      int size = 0;
+      for (int i = 0; i < elem_geoms.Size(); ++i)
+      {
+         fespace->GetLocalDerefinementMatrices(elem_geoms[i],
+                                               localR[elem_geoms[i]]);
+         geom_offsets[elem_geoms[i]] = size;
+         size += localR[elem_geoms[i]].TotalSize();
+      }
+      block_storage.SetSize(size);
+      bs_ptr = block_storage.HostWrite();
+      // copy blocks into block_storage
+      for (int i = 0; i < elem_geoms.Size(); ++i)
+      {
+         std::copy(localR[elem_geoms[i]].Data(),
+                   localR[elem_geoms[i]].Data()
+                   + localR[elem_geoms[i]].TotalSize(),
+                   bs_ptr);
+         bs_ptr += localR[elem_geoms[i]].TotalSize();
+      }
+   }
+
+   // second pass:
+   // - initialize buffers
+
+   {
+      auto ptr = send_permutations.HostWrite();
+      auto ptr2 = send_segment_idcs.HostWrite();
+      int i = 0;
+      for (auto &v : to_send)
+      {
+         ptr = std::copy(v.second.begin(), v.second.end(), ptr);
+         for (size_t idx = 0; idx < v.second.size(); ++idx)
+         {
+            *ptr2 = i;
+            ++ptr2;
+         }
+         ++i;
+      }
+   }
+
+   block_row_idcs_offsets[0] = 0;
+   block_col_idcs_offsets[0] = 0;
+   block_off_diag_row_idcs_offsets[0] = 0;
+   Array<int> mark(fespace->GetNDofs());
+   mark = 0;
+   {
+      int idx = 0;
+      recv_segments[0] = 0;
+      for (auto &v : od_seg_lens)
+      {
+         recv_ranks[idx] = v.first;
+         recv_segments[idx + 1] = recv_segments[idx] + v.second;
+         ++idx;
+      }
+   }
+   // key: index into dtrans.embeddings
+   // value: off-diagonal block offset, od_ridx, seg id
+   std::unordered_map<int, std::array<int, 3>> ks_map;
+   {
+      int od_ridx = 0;
+      int seg_id = 0;
+      for (auto &v1 : od_ks)
+      {
+         for (auto k : v1.second)
+         {
+            auto &tmp = ks_map[k];
+            tmp[0] = ks_map.size() - 1;
+            tmp[1] = od_ridx;
+            tmp[2] = seg_id;
+            od_ridx += get_ldofs(k);
+         }
+         ++seg_id;
+      }
+   }
+   int diag_idx = 0;
+   int var_offset = 0;
+   int ridx = 0;
+   int cidx = 0;
+   // can't break this up into separate diagonals/off-diagonals loops because
+   // of mark
+   for (int k = 0; k < dtrans.embeddings.Size(); ++k)
+   {
+      const Embedding &emb = dtrans.embeddings[k];
+      if (emb.parent < 0)
+      {
+         continue;
+      }
+      int fine_rank = old_ranks[k];
+      int coarse_rank = (emb.parent < 0) ? (-1 - emb.parent)
+                        : old_pncmesh->ElementRank(emb.parent);
+      if (coarse_rank == MyRank)
+      {
+         // either diagonal or off-diagonal
+         Geometry::Type geom =
+            fespace->GetMesh()->GetElementBaseGeometry(emb.parent);
+         if (fespace->IsVariableOrder())
+         {
+            const FiniteElement *fe = fespace->GetFE(emb.parent);
+            const DenseTensor &pmats = dtrans.point_matrices[geom];
+            const int ldof = fe->GetDof();
+
+            IsoparametricTransformation isotr;
+            isotr.SetIdentityTransformation(geom);
+
+            localRVO.SetSize(ldof, ldof);
+            isotr.SetPointMat(pmats(emb.matrix));
+            // Local restriction is size ldofxldof assuming that the parent
+            // and child are of same polynomial order.
+            fe->GetLocalRestriction(isotr, localRVO);
+            // copy block
+            auto s = localRVO.Height() * localRVO.Width();
+            std::copy(localRVO.Data(), localRVO.Data() + s, bs_ptr);
+            bs_ptr += s;
+         }
+         DenseMatrix &lR =
+            fespace->IsVariableOrder() ? localRVO : localR[geom](emb.matrix);
+         max_rows = std::max(lR.Height(), max_rows);
+         auto size = lR.Height() * lR.Width();
+         fespace->elem_dof->GetRow(emb.parent, dofs);
+         if (fine_rank == MyRank)
+         {
+            // diagonal
+            old_elem_dof->GetRow(k, old_dofs);
+            MFEM_VERIFY(old_dofs.Size() == dofs.Size(),
+                        "Parent and child must have same #dofs.");
+            block_row_idcs_offsets[diag_idx + 1] =
+               block_row_idcs_offsets[diag_idx] + lR.Height();
+            block_col_idcs_offsets[diag_idx + 1] =
+               block_col_idcs_offsets[diag_idx] + lR.Width();
+
+            if (fespace->IsVariableOrder())
+            {
+               block_offsets[diag_idx] = var_offset;
+               var_offset += size;
+            }
+            else
+            {
+               block_offsets[diag_idx] = geom_offsets[geom] + size * emb.matrix;
+            }
+            for (int i = 0; i < lR.Height(); ++i, ++ridx)
+            {
+               if (!std::isfinite(lR(i, 0)))
+               {
+                  row_idcs[ridx] = INT_MAX;
+                  continue;
+               }
+               int r = dofs[i];
+               int m = (r >= 0) ? r : (-1 - r);
+               if (is_dg || !mark[m])
+               {
+                  row_idcs[ridx] = r;
+                  mark[m] = 1;
+               }
+               else
+               {
+                  row_idcs[ridx] = INT_MAX;
+               }
+            }
+            for (int i = 0; i < lR.Width(); ++i, ++cidx)
+            {
+               col_idcs[cidx] = old_dofs[i];
+            }
+            ++diag_idx;
+         }
+         else
+         {
+            // off-diagonal
+            auto &tmp = ks_map.at(k);
+            auto od_idx = tmp[0];
+            auto od_ridx = tmp[1];
+            block_off_diag_row_idcs_offsets[od_idx + 1] =
+               block_off_diag_row_idcs_offsets[od_idx] + lR.Height();
+            block_off_diag_col_offsets[od_idx] = od_ridx;
+            block_off_diag_widths[od_idx] = lR.Width();
+            recv_segment_idcs[od_idx] = tmp[2];
+
+            if (fespace->IsVariableOrder())
+            {
+               off_diag_block_offsets[od_idx] = var_offset;
+               var_offset += size;
+            }
+            else
+            {
+               off_diag_block_offsets[od_idx] =
+                  geom_offsets[geom] + size * emb.matrix;
+            }
+            for (int i = 0; i < lR.Height(); ++i, ++od_ridx)
+            {
+               if (!std::isfinite(lR(i, 0)))
+               {
+                  row_off_diag_idcs[od_ridx] = INT_MAX;
+                  continue;
+               }
+               int r = dofs[i];
+               int m = (r >= 0) ? r : (-1 - r);
+               if (is_dg || !mark[m])
+               {
+                  row_off_diag_idcs[od_ridx] = r;
+                  mark[m] = 1;
+               }
+               else
+               {
+                  row_off_diag_idcs[od_ridx] = INT_MAX;
+               }
+            }
+            ++od_idx;
+         }
+      }
+   }
+   // if not using GPU, set max_rows/max_cols to zero
+   if (Device::Allows(Backend::DEVICE_MASK))
+   {
+      max_rows = std::min(max_rows, max_team_size);
+   }
+   else
+   {
+      max_rows = 1;
+   }
+   requests.reserve(recv_ranks.Size() + send_ranks.Size());
+}
+} // namespace mfem
+/// \endcond DO_NOT_DOCUMENT
+
+#endif

fem/pderefmat_op.hpp

@@ -0,0 +1,111 @@
+// 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.
+
+#ifndef MFEM_PDEREFMAT_OP
+#define MFEM_PDEREFMAT_OP
+
+#include "../config/config.hpp"
+
+#ifdef MFEM_USE_MPI
+
+#include "pfespace.hpp"
+
+#include "kernel_dispatch.hpp"
+
+#include <vector>
+
+namespace mfem
+{
+
+/// \cond DO_NOT_DOCUMENT
+struct ParDerefineMatrixOp : public Operator
+{
+   ParFiniteElementSpace *fespace;
+   /// offsets into block_storage for diagonal
+   Array<int> block_offsets;
+   /// offsets into row_idcs for diagonal
+   Array<int> block_row_idcs_offsets;
+   /// offsets into col_idcs for diagonal
+   Array<int> block_col_idcs_offsets;
+
+   /// offsets into block_storage for off-diagonal
+   Array<int> off_diag_block_offsets;
+   /// offsets into row_idcs for off-diagonal
+   Array<int> block_off_diag_row_idcs_offsets;
+   Array<int> block_off_diag_col_offsets;
+   Array<int> block_off_diag_widths;
+   /// mapping for row dofs, INT_MAX indicates the block row should be ignored.
+   /// negative means the row data should be negated.
+   /// only for diagonal blocks
+   Array<int> row_idcs;
+   /// mapping for col dofs, negative means the col data should be negated.
+   /// only for diagonal blocks
+   Array<int> col_idcs;
+
+   Array<int> pack_col_idcs;
+
+   /// mapping for row dofs, INT_MAX indicates the block row should be ignored.
+   /// negative means the row data should be negated.
+   /// only for off-diagonal blocks
+   Array<int> row_off_diag_idcs;
+   /// dense block matrices which can be reused to construct the full matrix
+   /// operation. These are stored contiguously and blocks have no restrictions
+   /// on shape (can be rectangle and differ from block to block).
+   /// This is only for the diagonal block.
+   Vector block_storage;
+   /// maximum height of any block in block_storage for GPU
+   /// parallelization, or 1 for CPU runs.
+   int max_rows;
+
+   /// quasi Ordering::byNODES, broken into sections by ranks we need to send
+   /// the data to
+   mutable Vector xghost_send;
+   /// quasi Ordering::byNODES, broken into sections by ranks we received
+   /// the data from
+   mutable Vector xghost_recv;
+   /// maps off-diagonal k to segment
+   Array<int> recv_segment_idcs;
+   /// cumulative count of dofs which will be received from other ranks
+   Array<int> recv_segments;
+   /// Source rank of each recv segment
+   Array<int> recv_ranks;
+   /// What send segment each entry in send_permutations corresponds to
+   Array<int> send_segment_idcs;
+   /// cumulative count of dofs which will be sent to other ranks
+   Array<int> send_segments;
+   /// Destination rank of each send segment
+   Array<int> send_ranks;
+   /// how to permute/sign change values from our local x to send to other ranks
+   Array<int> send_permutations;
+   /// internal buffer for MPI requests
+   mutable std::vector<MPI_Request> requests;
+
+   using MultKernelType = void (*)(const ParDerefineMatrixOp &, const Vector &,
+                                   Vector &);
+   /// template args: ordering, atomic
+   MFEM_REGISTER_KERNELS(MultKernel, MultKernelType, (Ordering::Type, bool));
+
+   struct Kernels
+   {
+      Kernels();
+   };
+
+   void Mult(const Vector &x, Vector &y) const;
+
+   ParDerefineMatrixOp(ParFiniteElementSpace &fespace_, int old_ndofs,
+                       const Table *old_elem_dof, const Table *old_elem_fos);
+};
+/// \endcond DO_NOT_DOCUMENT
+} // namespace mfem
+
+#endif
+
+#endif

fem/pfespace.cpp

@@ -22,12 +22,13 @@
 #include "../mesh/mesh_headers.hpp"
 #include "../general/binaryio.hpp"
 
+#include "pderefmat_op.hpp"
+
 #include <limits>
 #include <list>
 
 namespace mfem
 {
-
 ParFiniteElementSpace::ParFiniteElementSpace(
    const ParFiniteElementSpace &orig, ParMesh *pmesh,
    const FiniteElementCollection *fec)
@@ -4487,13 +4488,6 @@ ParFiniteElementSpace::RebalanceMatrix(int old_ndofs,
    return M;
 }
 
-
-struct DerefDofMessage
-{
-   std::vector<HYPRE_BigInt> dofs;
-   MPI_Request request;
-};
-
 HypreParMatrix*
 ParFiniteElementSpace::ParallelDerefinementMatrix(int old_ndofs,
                                                   const Table* old_elem_dof,
@@ -4536,7 +4530,13 @@ ParFiniteElementSpace::ParallelDerefinementMatrix(int old_ndofs,
       old_pncmesh->GetDerefinementTransforms();
    const Array<int> &old_ranks = old_pncmesh->GetDerefineOldRanks();
 
-   std::map<int, DerefDofMessage> messages;
+   // key: other rank
+   // value: send or recieve buffer
+   std::map<int, std::vector<HYPRE_BigInt>> to_send;
+   std::map<int, std::vector<HYPRE_BigInt>> to_recv;
+   // key: index into dtrans.embeddings
+   // value: [start, stop]
+   std::unordered_map<int, std::array<size_t, 2>> recv_messages;
 
    HYPRE_BigInt old_offset = HYPRE_AssumedPartitionCheck()
                              ? old_dof_offsets[0] : old_dof_offsets[MyRank];
@@ -4556,30 +4556,46 @@ ParFiniteElementSpace::ParallelDerefinementMatrix(int old_ndofs,
          old_elem_dof->GetRow(k, dofs);
          DofsToVDofs(dofs, old_ndofs);
 
-         DerefDofMessage &msg = messages[k];
-         msg.dofs.resize(dofs.Size());
+         std::vector<HYPRE_BigInt>& send_buf = to_send[coarse_rank];
+         auto pos = send_buf.size();
+         send_buf.resize(pos + dofs.Size());
          for (int i = 0; i < dofs.Size(); i++)
          {
-            msg.dofs[i] = old_offset + dofs[i];
+            send_buf[pos + i] = old_offset + dofs[i];
          }
-
-         MPI_Isend(&msg.dofs[0], static_cast<int>(msg.dofs.size()), HYPRE_MPI_BIG_INT,
-                   coarse_rank, 291, MyComm, &msg.request);
       }
       else if (coarse_rank == MyRank && fine_rank != MyRank)
       {
          MFEM_ASSERT(emb.parent >= 0, "");
          Geometry::Type geom = mesh->GetElementBaseGeometry(emb.parent);
 
-         DerefDofMessage &msg = messages[k];
-         msg.dofs.resize(ldof[geom]*vdim);
-
-         MPI_Irecv(&msg.dofs[0], ldof[geom]*vdim, HYPRE_MPI_BIG_INT,
-                   fine_rank, 291, MyComm, &msg.request);
+         std::vector<HYPRE_BigInt>& recv_buf = to_recv[fine_rank];
+         auto& msg = recv_messages[k];
+         msg[0] = recv_buf.size();
+         recv_buf.resize(recv_buf.size() + ldof[geom] * vdim);
+         msg[1] = recv_buf.size();
       }
-      // TODO: coalesce Isends/Irecvs to the same rank. Typically, on uniform
-      // derefinement, there should be just one send to MyRank-1 and one recv
-      // from MyRank+1
+   }
+
+   // assume embedding orders are consistent (i.e. what we expect to receive
+   // first from a given rank is sent first, etc.)
+   std::vector<MPI_Request> requests;
+   requests.reserve(to_send.size() + to_recv.size());
+   // enqueue recvs
+   for (auto &v : to_recv)
+   {
+      requests.emplace_back();
+      MPI_Irecv(v.second.data(), v.second.size(), HYPRE_MPI_BIG_INT, v.first,
+                MessageTag::DEREFINEMENT_MATRIX_CONSTRUCTION_DATA, MyComm,
+                &requests.back());
+   }
+   // enqueue sends
+   for (auto &v : to_send)
+   {
+      requests.emplace_back();
+      MPI_Isend(v.second.data(), v.second.size(), HYPRE_MPI_BIG_INT, v.first,
+                MessageTag::DEREFINEMENT_MATRIX_CONSTRUCTION_DATA, MyComm,
+                &requests.back());
    }
 
    DenseTensor localR[Geometry::NumGeom];
@@ -4637,10 +4653,7 @@ ParFiniteElementSpace::ParallelDerefinementMatrix(int old_ndofs,
    diag->Finalize();
 
    // wait for all sends/receives to complete
-   for (auto it = messages.begin(); it != messages.end(); ++it)
-   {
-      MPI_Wait(&it->second.request, MPI_STATUS_IGNORE);
-   }
+   MPI_Waitall(requests.size(), requests.data(), MPI_STATUSES_IGNORE);
 
    // create the off-diagonal part of the derefinement matrix
    SparseMatrix *offd = new SparseMatrix(ndofs*vdim, 1);
@@ -4661,13 +4674,14 @@ ParFiniteElementSpace::ParallelDerefinementMatrix(int old_ndofs,
 
          elem_dof->GetRow(emb.parent, dofs);
 
-         DerefDofMessage &msg = messages[k];
-         MFEM_ASSERT(msg.dofs.size(), "");
+         auto& odofs = to_recv.at(fine_rank);
+         auto &msg = recv_messages[k];
+         MFEM_ASSERT(msg[1] > msg[0], "");
 
          for (int vd = 0; vd < vdim; vd++)
          {
             MFEM_ASSERT(ldof[geom], "");
-            HYPRE_BigInt* remote_dofs = &msg.dofs[vd*ldof[geom]];
+            HYPRE_BigInt *remote_dofs = odofs.data() + msg[0] + vd * ldof[geom];
 
             for (int i = 0; i < lR.Height(); i++)
             {
@@ -4694,7 +4708,6 @@ ParFiniteElementSpace::ParallelDerefinementMatrix(int old_ndofs,
       }
    }
 
-   messages.clear();
    offd->Finalize(0);
    offd->SetWidth(static_cast<int>(col_map.size()));
 
@@ -4946,8 +4959,13 @@ void ParFiniteElementSpace::Update(bool want_transform)
 
          case Mesh::DEREFINE:
          {
+#if 0
             Th.Reset(ParallelDerefinementMatrix(old_ndofs, old_elem_dof,
                                                 old_elem_fos));
+#else
+            Th.Reset(new ParDerefineMatrixOp(*this, old_ndofs, old_elem_dof,
+                                             old_elem_fos));
+#endif
             if (Nonconforming())
             {
                Th.SetOperatorOwner(false);

fem/pfespace.hpp

@@ -24,9 +24,12 @@
 namespace mfem
 {
 
+struct ParDerefineMatrixOp;
+
 /// Abstract parallel finite element space.
 class ParFiniteElementSpace : public FiniteElementSpace
 {
+   friend struct ParDerefineMatrixOp;
 private:
    /// MPI data.
    MPI_Comm MyComm;

general/communication.hpp

@@ -425,16 +425,24 @@ public:
    ~GroupCommunicator();
 };
 
+/// General MPI message tags used by MFEM
+enum MessageTag
+{
+   DEREFINEMENT_MATRIX_CONSTRUCTION_DATA =
+      291, /// ParFiniteElementSpace ParallelDerefinementMatrix and
+   /// ParDerefineMatrixOp
+};
+
 enum VarMessageTag
 {
-   NEIGHBOR_ELEMENT_RANK_VM,  ///< NeighborElementRankMessage
-   NEIGHBOR_ORDER_VM,         ///< NeighborOrderMessage
-   NEIGHBOR_DEREFINEMENT_VM,  ///< NeighborDerefinementMessage
-   NEIGHBOR_REFINEMENT_VM,    ///< NeighborRefinementMessage
-   NEIGHBOR_PREFINEMENT_VM,   ///< NeighborPRefinementMessage
-   NEIGHBOR_ROW_VM,           ///< NeighborRowMessage
-   REBALANCE_VM,              ///< RebalanceMessage
-   REBALANCE_DOF_VM           ///< RebalanceDofMessage
+   NEIGHBOR_ELEMENT_RANK_VM, ///< NeighborElementRankMessage
+   NEIGHBOR_ORDER_VM,        ///< NeighborOrderMessage
+   NEIGHBOR_DEREFINEMENT_VM, ///< NeighborDerefinementMessage
+   NEIGHBOR_REFINEMENT_VM,   ///< NeighborRefinementMessage
+   NEIGHBOR_PREFINEMENT_VM,  ///< NeighborPRefinementMessage
+   NEIGHBOR_ROW_VM,          ///< NeighborRowMessage
+   REBALANCE_VM,             ///< RebalanceMessage
+   REBALANCE_DOF_VM,         ///< RebalanceDofMessage
 };
 
 /// \brief Variable-length MPI message containing unspecific binary data.

general/device.cpp

@@ -615,7 +615,7 @@ void Device::Setup(const std::string &device_option, const int device_id)
    if (Allows(Backend::DEBUG_DEVICE)) { ngpu = 1; }
 }
 
-MemoryType Device::QueryMemoryType(void *ptr)
+MemoryType Device::QueryMemoryType(const void* ptr)
 {
    // from HYPRE's hypre_GetPointerLocation
    MemoryType res = MemoryType::HOST;

general/device.hpp

@@ -297,9 +297,9 @@ public:
    /// Get the status of GPU-aware MPI flag.
    static bool GetGPUAwareMPI() { return Get().mpi_gpu_aware; }
 
-   /** @brief Query the device driver for what memory type a given @a ptr is
-       allocated with. */
-   static MemoryType QueryMemoryType(void *ptr);
+   /** Query the device driver for what memory type a given @a ptr is allocated
+    * with. */
+   static MemoryType QueryMemoryType(const void* ptr);
 
    /** @brief The number of hardware compute units/streaming multiprocessors
        available on a given compute device @a device_id. */

mesh/mesh.cpp

@@ -13257,7 +13257,7 @@ void Mesh::ScaleElements(real_t sf)
    delete [] vn;
 }
 
-void Mesh::Transform(void (*f)(const Vector&, Vector&))
+void Mesh::Transform(std::function<void(const Vector &, Vector&)> f)
 {
    // TODO: support for different new spaceDim.
    if (Nodes == NULL)
@@ -13270,7 +13270,7 @@ void Mesh::Transform(void (*f)(const Vector&, Vector&))
             vold(j) = vertices[i](j);
          }
          vnew.SetData(vertices[i]());
-         (*f)(vold, vnew);
+         f(vold, vnew);
       }
    }
    else

mesh/mesh.hpp

@@ -2254,7 +2254,7 @@ public:
    void ScaleSubdomains (real_t sf);
    void ScaleElements (real_t sf);
 
-   void Transform(void (*f)(const Vector&, Vector&));
+   void Transform(std::function<void(const Vector &, Vector&)> f);
    void Transform(VectorCoefficient &deformation);
 
    /** @brief This function should be called after the mesh node coordinates

tests/unit/fem/test_derefine.cpp

@@ -235,7 +235,7 @@ void test_derefine_L2_element(int order, Element::Type el_type, int basis_type)
    }
 }
 
-TEST_CASE("AMR Coarsen L2 Element","[AMR][Coarsen]")
+TEST_CASE("AMR Coarsen L2 Element","[AMR][Coarsen][CUDA]")
 {
    std::vector<int> orders_1d{0,1,2,3};
    std::vector<int> orders_2d{0,1,2,3};
@@ -420,7 +420,7 @@ void stress_parallel_coarsen(int order, Element::Type el_type, int basis_type)
 #endif
 }
 
-TEST_CASE("Parallel AMR Coarsen Stress Test", "[AMR][Coarsen][Parallel]")
+TEST_CASE("Parallel AMR Coarsen Stress Test", "[AMR][Coarsen][Parallel][CUDA]")
 {
    std::vector<int> orders_1d{0,1,2,3};
    std::vector<int> orders_2d{0,1,2,3};
@@ -482,7 +482,7 @@ TEST_CASE("Parallel AMR Coarsen Stress Test", "[AMR][Coarsen][Parallel]")
 
 }
 
-TEST_CASE("ParDerefine", "[Parallel]")
+TEST_CASE("ParDerefine", "[Parallel][CUDA]")
 {
    for (dimension = 2; dimension <= 3; ++dimension)
    {