1# libCEED: the CEED API Library 2 3[](https://travis-ci.org/CEED/libCEED) 4[](https://codecov.io/gh/CEED/libCEED/) 5[](https://opensource.org/licenses/BSD-2-Clause) 6[](https://codedocs.xyz/CEED/libCEED/) 7 8## Code for Efficient Extensible Discretization 9 10This repository contains an initial low-level API library for the efficient 11high-order discretization methods developed by the ECP co-design [Center for 12Efficient Exascale Discretizations (CEED)](http://ceed.exascaleproject.org). 13While our focus is on high-order finite elements, the approach is mostly 14algebraic and thus applicable to other discretizations in factored form, as 15explained in the API documentation portion of the [Doxygen documentation](https://codedocs.xyz/CEED/libCEED/md_doc_libCEEDapi.html). 16 17One of the challenges with high-order methods is that a global sparse matrix is 18no longer a good representation of a high-order linear operator, both with 19respect to the FLOPs needed for its evaluation, as well as the memory transfer 20needed for a matvec. Thus, high-order methods require a new "format" that still 21represents a linear (or more generally non-linear) operator, but not through a 22sparse matrix. 23 24The goal of libCEED is to propose such a format, as well as supporting 25implementations and data structures, that enable efficient operator evaluation 26on a variety of computational device types (CPUs, GPUs, etc.). This new operator 27description is based on algebraically [factored form](https://codedocs.xyz/CEED/libCEED/md_doc_libCEEDapi.html), 28which is easy to incorporate in a wide variety of applications, without significant 29refactoring of their own discretization infrastructure. 30 31The repository is part of the [CEED software suite][ceed-soft], a collection of 32software benchmarks, miniapps, libraries and APIs for efficient exascale 33discretizations based on high-order finite element and spectral element methods. 34See http://github.com/ceed for more information and source code availability. 35 36The CEED research is supported by the [Exascale Computing Project][ecp] 37(17-SC-20-SC), a collaborative effort of two U.S. Department of Energy 38organizations (Office of Science and the National Nuclear Security 39Administration) responsible for the planning and preparation of a [capable 40exascale ecosystem](https://exascaleproject.org/what-is-exascale), including 41software, applications, hardware, advanced system engineering and early testbed 42platforms, in support of the nation’s exascale computing imperative. 43 44For more details on the CEED API see http://ceed.exascaleproject.org/ceed-code/. 45 46## Building 47 48The CEED library, `libceed`, is a C99 library with no external dependencies. It 49can be built using 50 51 make 52 53or, with optimization flags 54 55 make OPT='-O3 -march=skylake-avx512 -ffp-contract=fast' 56 57These optimization flags are used by all languages (C, C++, Fortran) and this 58makefile variable can also be set for testing and examples (below). 59 60The library attempts to automatically detect support for the AVX 61instruction set using gcc-style compiler options for the host. 62Support may need to be manually specified via 63 64 make AVX=1 65 66or 67 68 make AVX=0 69 70if your compiler does not support gcc-style options, if you are cross 71compiling, etc. 72 73## Testing 74 75The test suite produces [TAP](https://testanything.org) output and is run by: 76 77 make test 78 79or, using the `prove` tool distributed with Perl (recommended) 80 81 make prove 82 83## Backends 84 85There are multiple supported backends, which can be selected at runtime in the examples: 86 87| CEED resource | Backend | 88| :----------------------- | :------------------------------------------------ | 89| `/cpu/self/ref/serial` | Serial reference implementation | 90| `/cpu/self/ref/blocked` | Blocked refrence implementation | 91| `/cpu/self/tmpl` | Backend template, delegates to `/cpu/self/ref/blocked` | 92| `/cpu/self/avx/serial` | Serial AVX implementation | 93| `/cpu/self/avx/blocked` | Blocked AVX implementation | 94| `/cpu/self/xsmm/serial` | Serial LIBXSMM implementation | 95| `/cpu/self/xsmm/blocked` | Blocked LIBXSMM implementation | 96| `/cpu/occa` | Serial OCCA kernels | 97| `/gpu/occa` | CUDA OCCA kernels | 98| `/omp/occa` | OpenMP OCCA kernels | 99| `/ocl/occa` | OpenCL OCCA kernels | 100| `/gpu/cuda/ref` | Reference pure CUDA kernels | 101| `/gpu/cuda/reg` | Pure CUDA kernels using one thread per element | 102| `/gpu/magma` | CUDA MAGMA kernels | 103 104 105The `/cpu/self/*/serial` backends process one element at a time and are intended for meshes 106with a smaller number of high order elements. The `/cpu/self/*/blocked` backends process 107blocked batches of eight interlaced elements and are intended for meshes with higher numbers 108of elements. 109 110The `/cpu/self/ref/*` backends are written in pure C and provide basic functionality. 111 112The `/cpu/self/avx/*` backends rely upon AVX instructions to provide vectorized CPU performance. 113 114The `/cpu/self/xsmm/*` backends rely upon the [LIBXSMM](http://github.com/hfp/libxsmm) package 115to provide vectorized CPU performance. The LIBXSMM backend does not use BLAS or MKL; however, 116if LIBXSMM was linked to MKL, this can be specified with the compilation flag `MKL=1`. 117 118The `/*/occa` backends rely upon the [OCCA](http://github.com/libocca/occa) package to provide 119cross platform performance. 120 121The `/gpu/cuda/*` backends provide GPU performance strictly using CUDA. 122 123The `/gpu/magma` backend relies upon the [MAGMA](https://bitbucket.org/icl/magma) package. 124 125## Examples 126 127libCEED comes with several examples of its usage, ranging from standalone C 128codes in the `/examples/ceed` directory to examples based on external packages, 129such as MFEM, PETSc, and Nek5000. Nek5000 v18.0 or greater is required. 130 131To build the examples, set the `MFEM_DIR`, `PETSC_DIR` and `NEK5K_DIR` variables 132and run: 133 134```console 135# libCEED examples on CPU and GPU 136cd examples/ceed 137make 138./ex1 -ceed /cpu/self 139./ex1 -ceed /gpu/occa 140cd ../.. 141 142# MFEM+libCEED examples on CPU and GPU 143cd examples/mfem 144make 145./bp1 -ceed /cpu/self -no-vis 146./bp1 -ceed /gpu/occa -no-vis 147cd ../.. 148 149# PETSc+libCEED examples on CPU and GPU 150cd examples/petsc 151make 152./bp1 -ceed /cpu/self 153./bp1 -ceed /gpu/occa 154cd ../.. 155 156cd navier-stokes 157make 158./navierstokes -ceed /cpu/self 159./navierstokes -ceed /gpu/occa 160cd ../.. 161 162# Nek+libCEED examples on CPU and GPU 163cd examples/nek5000 164./make-nek-examples.sh 165./run-nek-example.sh -ceed /cpu/self -b 3 166./run-nek-example.sh -ceed /gpu/occa -b 3 167cd ../.. 168``` 169 170The above code assumes a GPU-capable machine with the OCCA backend 171enabled. Depending on the available backends, other Ceed resource specifiers can 172be provided with the `-ceed` option. 173 174## Benchmarks 175 176A sequence of benchmarks for all enabled backends can be run using 177 178```console 179make benchmarks 180``` 181 182The results from the benchmarks are stored inside the `benchmarks/` directory 183and they can be viewed using the commands (requires python with matplotlib): 184 185```console 186cd benchmarks 187python postprocess-plot.py petsc-bp1-*-output.txt 188python postprocess-plot.py petsc-bp3-*-output.txt 189``` 190 191Using the `benchmarks` target runs a comprehensive set of benchmarks which may 192take some time to run. Subsets of the benchmarks can be run using targets such 193as `make bench-petsc-bp1`, or `make bench-petsc-bp3`. 194 195For more details about the benchmarks, see 196[`benchmarks/README.md`](benchmarks/README.md) 197 198 199## Install 200 201To install libCEED, run 202 203 make install prefix=/usr/local 204 205or (e.g., if creating packages), 206 207 make install prefix=/usr DESTDIR=/packaging/path 208 209Note that along with the library, libCEED installs kernel sources, e.g. OCCA 210kernels are installed in `$prefix/lib/okl`. This allows the OCCA backend to 211build specialized kernels at run-time. In a normal setting, the kernel sources 212will be found automatically (relative to the library file `libceed.so`). 213However, if that fails (e.g. if `libceed.so` is moved), one can copy (cache) the 214kernel sources inside the user OCCA directory, `~/.occa` using 215 216 $(OCCA_DIR)/bin/occa cache ceed $(CEED_DIR)/lib/okl/*.okl 217 218This will allow OCCA to find the sources regardless of the location of the CEED 219library. One may occasionally need to clear the OCCA cache, which can be accomplished 220by removing the `~/.occa` directory or by calling `$(OCCA_DIR)/bin/occa clear -a`. 221 222### pkg-config 223 224In addition to library and header, libCEED provides a [pkg-config][pkg-config1] 225file that can be used to easily compile and link. [For example][pkg-config2], if 226`$prefix` is a standard location or you set the environment variable 227`PKG_CONFIG_PATH`, 228 229 cc `pkg-config --cflags --libs ceed` -o myapp myapp.c 230 231will build `myapp` with libCEED. This can be used with the source or 232installed directories. Most build systems have support for pkg-config. 233 234## Contact 235 236You can reach the libCEED team by emailing [ceed-users@llnl.gov](mailto:ceed-users@llnl.gov) 237or by leaving a comment in the [issue tracker](https://github.com/CEED/libCEED/issues). 238 239## Copyright 240 241The following copyright applies to each file in the CEED software suite, unless 242otherwise stated in the file: 243 244> Copyright (c) 2017, Lawrence Livermore National Security, LLC. Produced at the 245> Lawrence Livermore National Laboratory. LLNL-CODE-734707. All Rights reserved. 246 247See files LICENSE and NOTICE for details. 248 249[ceed-soft]: http://ceed.exascaleproject.org/software/ 250[ecp]: https://exascaleproject.org/exascale-computing-project 251[pkg-config1]: https://en.wikipedia.org/wiki/Pkg-config 252[pkg-config2]: https://people.freedesktop.org/~dbn/pkg-config-guide.html#faq 253
){kind=link}
){kind=link}
){kind=link}
){kind=link}