Getting Started


The CEED library, libceed, is a C99 library with no required dependencies, and with Fortran, Python, Julia, and Rust interfaces. It can be built using:

$ make

or, with optimization flags:

$ make OPT='-O3 -march=skylake-avx512 -ffp-contract=fast'

These optimization flags are used by all languages (C, C++, Fortran) and this makefile variable can also be set for testing and examples (below).

The library attempts to automatically detect support for the AVX instruction set using gcc-style compiler options for the host. Support may need to be manually specified via:

$ make AVX=1


$ make AVX=0

if your compiler does not support gcc-style options, if you are cross compiling, etc.

To enable CUDA support, add CUDA_DIR=/opt/cuda or an appropriate directory to your make invocation. To enable HIP support, add ROCM_DIR=/opt/rocm or an appropriate directory. To enable SYCL support, add SYCL_DIR=/opt/sycl or an appropriate directory. Note that SYCL backends require building with oneAPI compilers as well:

$ . /opt/intel/oneapi/
$ make SYCL_DIR=/opt/intel/oneapi/compiler/latest/linux SYCLCXX=icpx CC=icx CXX=icpx

The library can be configured for host applications which use OpenMP paralellism via:

$ make OPENMP=1

which will allow operators created and applied from different threads inside an omp parallel region.

To store these or other arguments as defaults for future invocations of make, use:

$ make configure CUDA_DIR=/usr/local/cuda ROCM_DIR=/opt/rocm OPT='-O3 -march=znver2'

which stores these variables in


libCEED can be built for WASM using Emscripten. For example, one can build the library and run a standalone WASM executable using

$ emmake make build/ex2-surface.wasm
$ wasmer build/ex2-surface.wasm -- -s 200000

Additional Language Interfaces

The Fortran interface is built alongside the library automatically.

Python users can install using:

$ pip install libceed

or in a clone of the repository via pip install ..

Julia users can install using:

$ julia
julia> ]
pkg> add LibCEED

See the LibCEED.jl documentation for more information.

Rust users can include libCEED via Cargo.toml:

libceed = "0.12.0"

See the Cargo documentation for details.


The test suite produces TAP output and is run by:

$ make test

or, using the prove tool distributed with Perl (recommended):

$ make prove


There are multiple supported backends, which can be selected at runtime in the examples:

CEED resource


Deterministic Capable

CPU Native


Serial reference implementation



Blocked reference implementation



Serial optimized C implementation



Blocked optimized C implementation



Serial AVX implementation



Blocked AVX implementation


CPU Valgrind


Memcheck backends, undefined value checks




Serial LIBXSMM implementation



Blocked LIBXSMM implementation


CUDA Native


Reference pure CUDA kernels



Optimized pure CUDA kernels using shared memory



Optimized pure CUDA kernels using code generation


HIP Native


Reference pure HIP kernels



Optimized pure HIP kernels using shared memory



Optimized pure HIP kernels using code generation


SYCL Native


Reference pure SYCL kernels



Optimized pure SYCL kernels using shared memory




CUDA MAGMA kernels



CUDA MAGMA kernels



HIP MAGMA kernels



HIP MAGMA kernels




Selects backend based on available OCCA modes



OCCA backend with serial CPU kernels



OCCA backend with OpenMP kernels



OCCA backend with DPC++ kernels



OCCA backend with CUDA kernels



OCCA backend with HIP kernels


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/ref/* backends are written in pure C and provide basic functionality.

The /cpu/self/opt/* backends are written in pure C and use partial e-vectors to improve performance.

The /cpu/self/avx/* backends rely upon AVX instructions to provide vectorized CPU performance.

The /cpu/self/memcheck/* backends rely upon the Valgrind Memcheck tool to help verify that user QFunctions have no undefined values. To use, run your code with Valgrind and the Memcheck backends, e.g. valgrind ./build/ex1 -ceed /cpu/self/ref/memcheck. A ‘development’ or ‘debugging’ version of Valgrind with headers is required to use this backend. This backend can be run in serial or blocked mode and defaults to running in the serial mode if /cpu/self/memcheck is selected at runtime.

The /cpu/self/xsmm/* backends rely upon the LIBXSMM package to provide vectorized CPU performance. If linking MKL and LIBXSMM is desired but the Makefile is not detecting MKLROOT, linking libCEED against MKL can be forced by setting the environment variable MKL=1. The LIBXSMM main development branch from 7 April 2024 or newer is required.

The /gpu/cuda/* backends provide GPU performance strictly using CUDA.

The /gpu/hip/* backends provide GPU performance strictly using HIP. They are based on the /gpu/cuda/* backends. ROCm version 4.2 or newer is required.

The /gpu/sycl/* backends provide GPU performance strictly using SYCL. They are based on the /gpu/cuda/* and /gpu/hip/* backends.

The /gpu/*/magma/* backends rely upon the MAGMA package. To enable the MAGMA backends, the environment variable MAGMA_DIR must point to the top-level MAGMA directory, with the MAGMA library located in $(MAGMA_DIR)/lib/. By default, MAGMA_DIR is set to ../magma; to build the MAGMA backends with a MAGMA installation located elsewhere, create a link to magma/ in libCEED’s parent directory, or set MAGMA_DIR to the proper location. MAGMA version 2.5.0 or newer is required. Currently, each MAGMA library installation is only built for either CUDA or HIP. The corresponding set of libCEED backends (/gpu/cuda/magma/* or /gpu/hip/magma/*) will automatically be built for the version of the MAGMA library found in MAGMA_DIR.

Users can specify a device for all CUDA, HIP, and MAGMA backends through adding :device_id=# after the resource name. For example:

  • /gpu/cuda/gen:device_id=1

The /*/occa backends rely upon the OCCA package to provide cross platform performance. To enable the OCCA backend, the environment variable OCCA_DIR must point to the top-level OCCA directory, with the OCCA library located in the ${OCCA_DIR}/lib (By default, OCCA_DIR is set to ../occa). OCCA version 1.4.0 or newer is required.

Users can pass specific OCCA device properties after setting the CEED resource. For example:

  • "/*/occa:mode='CUDA',device_id=0"

Bit-for-bit reproducibility is important in some applications. However, some libCEED backends use non-deterministic operations, such as atomicAdd for increased performance. The backends which are capable of generating reproducible results, with the proper compilation options, are highlighted in the list above.


libCEED comes with several examples of its usage, ranging from standalone C codes in the /examples/ceed directory to examples based on external packages, such as MFEM, PETSc, and Nek5000. Nek5000 v18.0 or greater is required.

To build the examples, set the MFEM_DIR, PETSC_DIR (and optionally PETSC_ARCH), and NEK5K_DIR variables and run:

$ cd examples/
# libCEED examples on CPU and GPU
$ cd ceed/
$ make
$ ./ex1-volume -ceed /cpu/self
$ ./ex1-volume -ceed /gpu/cuda
$ ./ex2-surface -ceed /cpu/self
$ ./ex2-surface -ceed /gpu/cuda
$ cd ..

# MFEM+libCEED examples on CPU and GPU
$ cd mfem/
$ make
$ ./bp1 -ceed /cpu/self -no-vis
$ ./bp3 -ceed /gpu/cuda -no-vis
$ cd ..

# Nek5000+libCEED examples on CPU and GPU
$ cd nek/
$ make
$ ./ -e bp1 -ceed /cpu/self -b 3
$ ./ -e bp3 -ceed /gpu/cuda -b 3
$ cd ..

# PETSc+libCEED examples on CPU and GPU
$ cd petsc/
$ make
$ ./bps -problem bp1 -ceed /cpu/self
$ ./bps -problem bp2 -ceed /gpu/cuda
$ ./bps -problem bp3 -ceed /cpu/self
$ ./bps -problem bp4 -ceed /gpu/cuda
$ ./bps -problem bp5 -ceed /cpu/self
$ ./bps -problem bp6 -ceed /gpu/cuda
$ cd ..

$ cd petsc/
$ make
$ ./bpsraw -problem bp1 -ceed /cpu/self
$ ./bpsraw -problem bp2 -ceed /gpu/cuda
$ ./bpsraw -problem bp3 -ceed /cpu/self
$ ./bpsraw -problem bp4 -ceed /gpu/cuda
$ ./bpsraw -problem bp5 -ceed /cpu/self
$ ./bpsraw -problem bp6 -ceed /gpu/cuda
$ cd ..

$ cd petsc/
$ make
$ ./bpssphere -problem bp1 -ceed /cpu/self
$ ./bpssphere -problem bp2 -ceed /gpu/cuda
$ ./bpssphere -problem bp3 -ceed /cpu/self
$ ./bpssphere -problem bp4 -ceed /gpu/cuda
$ ./bpssphere -problem bp5 -ceed /cpu/self
$ ./bpssphere -problem bp6 -ceed /gpu/cuda
$ cd ..

$ cd petsc/
$ make
$ ./area -problem cube -ceed /cpu/self -degree 3
$ ./area -problem cube -ceed /gpu/cuda -degree 3
$ ./area -problem sphere -ceed /cpu/self -degree 3 -dm_refine 2
$ ./area -problem sphere -ceed /gpu/cuda -degree 3 -dm_refine 2

$ cd fluids/
$ make
$ ./navierstokes -ceed /cpu/self -degree 1
$ ./navierstokes -ceed /gpu/cuda -degree 1
$ cd ..

$ cd solids/
$ make
$ ./elasticity -ceed /cpu/self -mesh [.exo file] -degree 2 -E 1 -nu 0.3 -problem Linear -forcing mms
$ ./elasticity -ceed /gpu/cuda -mesh [.exo file] -degree 2 -E 1 -nu 0.3 -problem Linear -forcing mms
$ cd ..

For the last example shown, sample meshes to be used in place of [.exo file] can be found at

The above code assumes a GPU-capable machine with the CUDA backends enabled. Depending on the available backends, other CEED resource specifiers can be provided with the -ceed option. Other command line arguments can be found in examples/petsc.


A sequence of benchmarks for all enabled backends can be run using:

$ make benchmarks

The results from the benchmarks are stored inside the benchmarks/ directory and they can be viewed using the commands (requires python with matplotlib):

$ cd benchmarks
$ python petsc-bps-bp1-*-output.txt
$ python petsc-bps-bp3-*-output.txt

Using the benchmarks target runs a comprehensive set of benchmarks which may take some time to run. Subsets of the benchmarks can be run using the scripts in the benchmarks folder.

For more details about the benchmarks, see the benchmarks/ file.


To install libCEED, run:

$ make install prefix=/path/to/install/dir

or (e.g., if creating packages):

$ make install prefix=/usr DESTDIR=/packaging/path

To build and install in separate steps, run:

$ make for_install=1 prefix=/path/to/install/dir
$ make install prefix=/path/to/install/dir

The usual variables like CC and CFLAGS are used, and optimization flags for all languages can be set using the likes of OPT='-O3 -march=native'. Use STATIC=1 to build static libraries (libceed.a).

To install libCEED for Python, run:

$ pip install libceed

with the desired setuptools options, such as --user.


In addition to library and header, libCEED provides a pkg-config file that can be used to easily compile and link. For example, if $prefix is a standard location or you set the environment variable PKG_CONFIG_PATH:

$ cc `pkg-config --cflags --libs ceed` -o myapp myapp.c

will build myapp with libCEED. This can be used with the source or installed directories. Most build systems have support for pkg-config.


You can reach the libCEED team by emailing or by leaving a comment in the issue tracker.

How to Cite

If you utilize libCEED please cite:

  author       = {Jed Brown and Ahmad Abdelfattah and Valeria Barra and Natalie Beams and Jean Sylvain Camier and Veselin Dobrev and Yohann Dudouit and Leila Ghaffari and Tzanio Kolev and David Medina and Will Pazner and Thilina Ratnayaka and Jeremy Thompson and Stan Tomov},
  title        = {{libCEED}: Fast algebra for high-order element-based discretizations},
  journal      = {Journal of Open Source Software},
  year         = {2021},
  publisher    = {The Open Journal},
  volume       = {6},
  number       = {63},
  pages        = {2945},
  doi          = {10.21105/joss.02945}

The archival copy of the libCEED user manual is maintained on Zenodo. To cite the user manual:

  author       = {Abdelfattah, Ahmad and
                  Barra, Valeria and
                  Beams, Natalie and
                  Brown, Jed and
                  Camier, Jean-Sylvain and
                  Dobrev, Veselin and
                  Dudouit, Yohann and
                  Ghaffari, Leila and
                  Grimberg, Sebastian and
                  Kolev, Tzanio and
                  Medina, David and
                  Pazner, Will and
                  Ratnayaka, Thilina and
                  Shakeri, Rezgar and
                  Thompson, Jeremy L and
                  Tomov, Stanimire and
                  Wright III, James},
  title        = {{libCEED} User Manual},
  month        = nov,
  year         = 2023,
  publisher    = {Zenodo},
  version      = {0.12.0},
  doi          = {10.5281/zenodo.10062388}

For libCEED’s Python interface please cite:

  author    = {{V}aleria {B}arra and {J}ed {B}rown and {J}eremy {T}hompson and {Y}ohann {D}udouit},
  title     = {{H}igh-performance operator evaluations with ease of use: lib{C}{E}{E}{D}'s {P}ython interface},
  booktitle = {{P}roceedings of the 19th {P}ython in {S}cience {C}onference},
  pages     = {85 - 90},
  year      = {2020},
  editor    = {{M}eghann {A}garwal and {C}hris {C}alloway and {D}illon {N}iederhut and {D}avid {S}hupe},
  doi       = {10.25080/Majora-342d178e-00c}

The BibTeX entries for these references can be found in the doc/bib/references.bib file.