Changes/Release Notes

On this page we provide a summary of the main API changes, new features and examples for each release of libCEED.

Current Main

The current main (formerly called master) branch contains bug fixes and additional features.

Interface changes

New features

  • New HIP MAGMA backends for hipMAGMA library users: /gpu/hip/magma and /gpu/hip/magma/det.

  • Julia and Rust interfaces added, providing a nearly 1-1 correspondence with the C interface, plus some convenience features.

  • New HIP backends for improved tensor basis performance: /gpu/hip/shared and /gpu/hip/gen.

  • Static libraries can be built with make STATIC=1 and the pkg-config file is installed accordingly.

Performance improvements

Examples

v0.7 (Sep 29, 2020)

Interface changes

  • Replace limited CeedInterlaceMode with more flexible component stride compstride in CeedElemRestriction constructors. As a result, the indices parameter has been replaced with offsets and the nnodes parameter has been replaced with lsize. These changes improve support for mixed finite element methods.

  • Replace various uses of Ceed*Get*Status with Ceed*Is* in the backend API to match common nomenclature.

  • Replace CeedOperatorAssembleLinearDiagonal with CeedOperatorLinearAssembleDiagonal() for clarity.

  • Linear Operators can be assembled as point-block diagonal matrices with CeedOperatorLinearAssemblePointBlockDiagonal(), provided in row-major form in a ncomp by ncomp block per node.

  • Diagonal assemble interface changed to accept a CeedVector instead of a pointer to a CeedVector to reduce memory movement when interfacing with calling code.

  • Added CeedOperatorLinearAssembleAddDiagonal() and CeedOperatorLinearAssembleAddPointBlockDiagonal() for improved future integration with codes such as MFEM that compose the action of CeedOperators external to libCEED.

  • Added CeedVectorTakeAray() to sync and remove libCEED read/write access to an allocated array and pass ownership of the array to the caller. This function is recommended over CeedVectorSyncArray() when the CeedVector has an array owned by the caller that was set by CeedVectorSetArray().

  • Added CeedQFunctionContext object to manage user QFunction context data and reduce copies between device and host memory.

  • Added CeedOperatorMultigridLevelCreate(), CeedOperatorMultigridLevelCreateTensorH1(), and CeedOperatorMultigridLevelCreateH1() to facilitate creation of multigrid prolongation, restriction, and coarse grid operators using a common quadrature space.

New features

  • New HIP backend: /gpu/hip/ref.

  • CeedQFunction support for user CUfunctions in some backends

Performance improvements

  • OCCA backend rebuilt to facilitate future performance enhancements.

  • Petsc BPs suite improved to reduce noise due to multiple calls to mpiexec.

Examples

Deprecated backends

  • The /gpu/cuda/reg backend has been removed, with its core features moved into /gpu/cuda/ref and /gpu/cuda/shared.

v0.6 (Mar 29, 2020)

libCEED v0.6 contains numerous new features and examples, as well as expanded documentation in this new website.

New features

  • New Python interface using CFFI provides a nearly 1-1 correspondence with the C interface, plus some convenience features. For instance, data stored in the CeedVector structure are available without copy as numpy.ndarray. Short tutorials are provided in Binder.

  • Linear QFunctions can be assembled as block-diagonal matrices (per quadrature point, CeedOperatorAssembleLinearQFunction()) or to evaluate the diagonal (CeedOperatorAssembleLinearDiagonal()). These operations are useful for preconditioning ingredients and are used in the libCEED’s multigrid examples.

  • The inverse of separable operators can be obtained using CeedOperatorCreateFDMElementInverse() and applied with CeedOperatorApply(). This is a useful preconditioning ingredient, especially for Laplacians and related operators.

  • New functions: CeedVectorNorm(), CeedOperatorApplyAdd(), CeedQFunctionView(), CeedOperatorView().

  • Make public accessors for various attributes to facilitate writing composable code.

  • New backend: /cpu/self/memcheck/serial.

  • QFunctions using variable-length array (VLA) pointer constructs can be used with CUDA backends. (Single source is coming soon for OCCA backends.)

  • Fix some missing edge cases in CUDA backend.

Performance Improvements

  • MAGMA backend performance optimization and non-tensor bases.

  • No-copy optimization in CeedOperatorApply().

Interface changes

  • Replace CeedElemRestrictionCreateIdentity and CeedElemRestrictionCreateBlocked with more flexible CeedElemRestrictionCreateStrided() and CeedElemRestrictionCreateBlockedStrided().

  • Add arguments to CeedQFunctionCreateIdentity().

  • Replace ambiguous uses of CeedTransposeMode for L-vector identification with CeedInterlaceMode. This is now an attribute of the CeedElemRestriction (see CeedElemRestrictionCreate()) and no longer passed as lmode arguments to CeedOperatorSetField() and CeedElemRestrictionApply().

Examples

libCEED-0.6 contains greatly expanded examples with new documentation. Notable additions include:

  • Standalone Ex2-Surface (examples/ceed/ex2-surface): compute the area of a domain in 1, 2, and 3 dimensions by applying a Laplacian.

  • PETSc Area (examples/petsc/area.c): computes surface area of domains (like the cube and sphere) by direct integration on a surface mesh; demonstrates geometric dimension different from topological dimension.

  • PETSc Bakeoff problems and generalizations:

    • examples/petsc/bpsraw.c (formerly bps.c): transparent CUDA support.

    • examples/petsc/bps.c (formerly bpsdmplex.c): performance improvements and transparent CUDA support.

    • Bakeoff problems on the cubed-sphere (examples/petsc/bpssphere.c): generalizations of all CEED BPs to the surface of the sphere; demonstrates geometric dimension different from topological dimension.

  • Multigrid (examples/petsc/multigrid.c): new p-multigrid solver with algebraic multigrid coarse solve.

  • Compressible Navier-Stokes mini-app (examples/fluids/navierstokes.c; formerly examples/navier-stokes): unstructured grid support (using PETSc’s DMPlex), implicit time integration, SU/SUPG stabilization, free-slip boundary conditions, and quasi-2D computational domain support.

  • Solid mechanics mini-app (examples/solids/elasticity.c): new solver for linear elasticity, small-strain hyperelasticity, and globalized finite-strain hyperelasticity using p-multigrid with algebraic multigrid coarse solve.

v0.5 (Sep 18, 2019)

For this release, several improvements were made. Two new CUDA backends were added to the family of backends, of which, the new cuda-gen backend achieves state-of-the-art performance using single-source CeedQFunction. From this release, users can define Q-Functions in a single source code independently of the targeted backend with the aid of a new macro CEED QFUNCTION to support JIT (Just-In-Time) and CPU compilation of the user provided CeedQFunction code. To allow a unified declaration, the CeedQFunction API has undergone a slight change: the QFunctionField parameter ncomp has been changed to size. This change requires setting the previous value of ncomp to ncomp*dim when adding a QFunctionField with eval mode CEED EVAL GRAD.

Additionally, new CPU backends were included in this release, such as the /cpu/self/opt/* backends (which are written in pure C and use partial E-vectors to improve performance) and the /cpu/self/ref/memcheck backend (which relies upon the Valgrind Memcheck tool to help verify that user CeedQFunction have no undefined values). This release also included various performance improvements, bug fixes, new examples, and improved tests. Among these improvements, vectorized instructions for CeedQFunction code compiled for CPU were enhanced by using CeedPragmaSIMD instead of CeedPragmaOMP, implementation of a CeedQFunction gallery and identity Q-Functions were introduced, and the PETSc benchmark problems were expanded to include unstructured meshes handling were. For this expansion, the prior version of the PETSc BPs, which only included data associated with structured geometries, were renamed bpsraw, and the new version of the BPs, which can handle data associated with any unstructured geometry, were called bps. Additionally, other benchmark problems, namely BP2 and BP4 (the vector-valued versions of BP1 and BP3, respectively), and BP5 and BP6 (the collocated versions—for which the quadrature points are the same as the Gauss Lobatto nodes—of BP3 and BP4 respectively) were added to the PETSc examples. Furthermoew, another standalone libCEED example, called ex2, which computes the surface area of a given mesh was added to this release.

Backends available in this release:

CEED resource (-ceed)

Backend

/cpu/self/ref/serial

Serial reference implementation

/cpu/self/ref/blocked

Blocked reference implementation

/cpu/self/ref/memcheck

Memcheck backend, undefined value checks

/cpu/self/opt/serial

Serial optimized C implementation

/cpu/self/opt/blocked

Blocked optimized C implementation

/cpu/self/avx/serial

Serial AVX implementation

/cpu/self/avx/blocked

Blocked AVX implementation

/cpu/self/xsmm/serial

Serial LIBXSMM implementation

/cpu/self/xsmm/blocked

Blocked LIBXSMM implementation

/cpu/occa

Serial OCCA kernels

/gpu/occa

CUDA OCCA kernels

/omp/occa

OpenMP OCCA kernels

/ocl/occa

OpenCL OCCA kernels

/gpu/cuda/ref

Reference pure CUDA kernels

/gpu/cuda/reg

Pure CUDA kernels using one thread per element

/gpu/cuda/shared

Optimized pure CUDA kernels using shared memory

/gpu/cuda/gen

Optimized pure CUDA kernels using code generation

/gpu/magma

CUDA MAGMA kernels

Examples available in this release:

User code

Example

ceed

  • ex1 (volume)

  • ex2 (surface)

mfem

  • BP1 (scalar mass operator)

  • BP3 (scalar Laplace operator)

petsc

  • BP1 (scalar mass operator)

  • BP2 (vector mass operator)

  • BP3 (scalar Laplace operator)

  • BP4 (vector Laplace operator)

  • BP5 (collocated scalar Laplace operator)

  • BP6 (collocated vector Laplace operator)

  • Navier-Stokes

nek5000

  • BP1 (scalar mass operator)

  • BP3 (scalar Laplace operator)

v0.4 (Apr 1, 2019)

libCEED v0.4 was made again publicly available in the second full CEED software distribution, release CEED 2.0. This release contained notable features, such as four new CPU backends, two new GPU backends, CPU backend optimizations, initial support for operator composition, performance benchmarking, and a Navier-Stokes demo. The new CPU backends in this release came in two families. The /cpu/self/*/serial backends process one element at a time and are intended for meshes with a smaller number of high order elements. The /cpu/self/*/blocked backends process blocked batches of eight interlaced elements and are intended for meshes with higher numbers of elements. The /cpu/self/avx/* backends rely upon AVX instructions to provide vectorized CPU performance. The /cpu/self/xsmm/* backends rely upon the LIBXSMM package to provide vectorized CPU performance. The /gpu/cuda/* backends provide GPU performance strictly using CUDA. The /gpu/cuda/ref backend is a reference CUDA backend, providing reasonable performance for most problem configurations. The /gpu/cuda/reg backend uses a simple parallelization approach, where each thread treats a finite element. Using just in time compilation, provided by nvrtc (NVidia Runtime Compiler), and runtime parameters, this backend unroll loops and map memory address to registers. The /gpu/cuda/reg backend achieve good peak performance for 1D, 2D, and low order 3D problems, but performance deteriorates very quickly when threads run out of registers.

A new explicit time-stepping Navier-Stokes solver was added to the family of libCEED examples in the examples/petsc directory (see Compressible Navier-Stokes mini-app). This example solves the time-dependent Navier-Stokes equations of compressible gas dynamics in a static Eulerian three-dimensional frame, using structured high-order finite/spectral element spatial discretizations and explicit high-order time-stepping (available in PETSc). Moreover, the Navier-Stokes example was developed using PETSc, so that the pointwise physics (defined at quadrature points) is separated from the parallelization and meshing concerns.

Backends available in this release:

CEED resource (-ceed)

Backend

/cpu/self/ref/serial

Serial reference implementation

/cpu/self/ref/blocked

Blocked reference implementation

/cpu/self/tmpl

Backend template, defaults to /cpu/self/blocked

/cpu/self/avx/serial

Serial AVX implementation

/cpu/self/avx/blocked

Blocked AVX implementation

/cpu/self/xsmm/serial

Serial LIBXSMM implementation

/cpu/self/xsmm/blocked

Blocked LIBXSMM implementation

/cpu/occa

Serial OCCA kernels

/gpu/occa

CUDA OCCA kernels

/omp/occa

OpenMP OCCA kernels

/ocl/occa

OpenCL OCCA kernels

/gpu/cuda/ref

Reference pure CUDA kernels

/gpu/cuda/reg

Pure CUDA kernels using one thread per element

/gpu/magma

CUDA MAGMA kernels

Examples available in this release:

User code

Example

ceed

ex1 (volume)

mfem

  • BP1 (scalar mass operator)

  • BP3 (scalar Laplace operator)

petsc

  • BP1 (scalar mass operator)

  • BP3 (scalar Laplace operator)

  • Navier-Stokes

nek5000

  • BP1 (scalar mass operator)

  • BP3 (scalar Laplace operator)

v0.3 (Sep 30, 2018)

Notable features in this release include active/passive field interface, support for non-tensor bases, backend optimization, and improved Fortran interface. This release also focused on providing improved continuous integration, and many new tests with code coverage reports of about 90%. This release also provided a significant change to the public interface: a CeedQFunction can take any number of named input and output arguments while CeedOperator connects them to the actual data, which may be supplied explicitly to CeedOperatorApply() (active) or separately via CeedOperatorSetField() (passive). This interface change enables reusable libraries of CeedQFunctions and composition of block solvers constructed using CeedOperator. A concept of blocked restriction was added to this release and used in an optimized CPU backend. Although this is typically not visible to the user, it enables effective use of arbitrary-length SIMD while maintaining cache locality. This CPU backend also implements an algebraic factorization of tensor product gradients to perform fewer operations than standard application of interpolation and differentiation from nodes to quadrature points. This algebraic formulation automatically supports non-polynomial and non-interpolatory bases, thus is more general than the more common derivation in terms of Lagrange polynomials on the quadrature points.

Backends available in this release:

CEED resource (-ceed)

Backend

/cpu/self/blocked

Blocked reference implementation

/cpu/self/ref

Serial reference implementation

/cpu/self/tmpl

Backend template, defaults to /cpu/self/blocked

/cpu/occa

Serial OCCA kernels

/gpu/occa

CUDA OCCA kernels

/omp/occa

OpenMP OCCA kernels

/ocl/occa

OpenCL OCCA kernels

/gpu/magma

CUDA MAGMA kernels

Examples available in this release:

User code

Example

ceed

ex1 (volume)

mfem

  • BP1 (scalar mass operator)

  • BP3 (scalar Laplace operator)

petsc

  • BP1 (scalar mass operator)

  • BP3 (scalar Laplace operator)

nek5000

  • BP1 (scalar mass operator)

  • BP3 (scalar Laplace operator)

v0.21 (Sep 30, 2018)

A MAGMA backend (which relies upon the MAGMA package) was integrated in libCEED for this release. This initial integration set up the framework of using MAGMA and provided the libCEED functionality through MAGMA kernels as one of libCEED’s computational backends. As any other backend, the MAGMA backend provides extended basic data structures for CeedVector, CeedElemRestriction, and CeedOperator, and implements the fundamental CEED building blocks to work with the new data structures. In general, the MAGMA-specific data structures keep the libCEED pointers to CPU data but also add corresponding device (e.g., GPU) pointers to the data. Coherency is handled internally, and thus seamlessly to the user, through the functions/methods that are provided to support them.

Backends available in this release:

CEED resource (-ceed)

Backend

/cpu/self

Serial reference implementation

/cpu/occa

Serial OCCA kernels

/gpu/occa

CUDA OCCA kernels

/omp/occa

OpenMP OCCA kernels

/ocl/occa

OpenCL OCCA kernels

/gpu/magma

CUDA MAGMA kernels

Examples available in this release:

User code

Example

ceed

ex1 (volume)

mfem

  • BP1 (scalar mass operator)

  • BP3 (scalar Laplace operator)

petsc

BP1 (scalar mass operator)

nek5000

BP1 (scalar mass operator)

v0.2 (Mar 30, 2018)

libCEED was made publicly available the first full CEED software distribution, release CEED 1.0. The distribution was made available using the Spack package manager to provide a common, easy-to-use build environment, where the user can build the CEED distribution with all dependencies. This release included a new Fortran interface for the library. This release also contained major improvements in the OCCA backend (including a new /ocl/occa backend) and new examples. The standalone libCEED example was modified to compute the volume volume of a given mesh (in 1D, 2D, or 3D) and placed in an examples/ceed subfolder. A new mfem example to perform BP3 (with the application of the Laplace operator) was also added to this release.

Backends available in this release:

CEED resource (-ceed)

Backend

/cpu/self

Serial reference implementation

/cpu/occa

Serial OCCA kernels

/gpu/occa

CUDA OCCA kernels

/omp/occa

OpenMP OCCA kernels

/ocl/occa

OpenCL OCCA kernels

Examples available in this release:

User code

Example

ceed

ex1 (volume)

mfem

  • BP1 (scalar mass operator)

  • BP3 (scalar Laplace operator)

petsc

BP1 (scalar mass operator)

nek5000

BP1 (scalar mass operator)

v0.1 (Jan 3, 2018)

Initial low-level API of the CEED project. The low-level API provides a set of Finite Elements kernels and components for writing new low-level kernels. Examples include: vector and sparse linear algebra, element matrix assembly over a batch of elements, partial assembly and action for efficient high-order operators like mass, diffusion, advection, etc. The main goal of the low-level API is to establish the basis for the high-level API. Also, identifying such low-level kernels and providing a reference implementation for them serves as the basis for specialized backend implementations. This release contained several backends: /cpu/self, and backends which rely upon the OCCA package, such as /cpu/occa, /gpu/occa, and /omp/occa. It also included several examples, in the examples folder: A standalone code that shows the usage of libCEED (with no external dependencies) to apply the Laplace operator, ex1; an mfem example to perform BP1 (with the application of the mass operator); and a petsc example to perform BP1 (with the application of the mass operator).

Backends available in this release:

CEED resource (-ceed)

Backend

/cpu/self

Serial reference implementation

/cpu/occa

Serial OCCA kernels

/gpu/occa

CUDA OCCA kernels

/omp/occa

OpenMP OCCA kernels

Examples available in this release:

User code

Example

ceed

ex1 (scalar Laplace operator)

mfem

BP1 (scalar mass operator)

petsc

BP1 (scalar mass operator)