OpenBLAS

GitHub Actions:

Azure Pipelines:

Introduction

OpenBLAS is an optimized BLAS (Basic Linear Algebra Subprograms) library based on GotoBLAS2 1.13 BSD version.

For more information about OpenBLAS, please see:

• The documentation at openmathlib.org/OpenBLAS/docs/,
• The home page at openmathlib.org/OpenBLAS/.

For a general introduction to the BLAS routines, please refer to the extensive documentation of their reference implementation hosted at netlib: https://www.netlib.org/blas. On that site you will likewise find documentation for the reference implementation of the higher-level library LAPACK - the Linear Algebra Package that comes included with OpenBLAS. If you are looking for a general primer or refresher on Linear Algebra, the set of six 20-minute lecture videos by Prof. Gilbert Strang on either MIT OpenCourseWare here or YouTube here may be helpful.

Binary Packages

We provide official binary packages for the following platform:

• Windows x86/x86_64
• Windows arm64 (woa)

You can download them from file hosting on sourceforge.net or from the Releases section of the GitHub project page.

OpenBLAS is also packaged for many package managers - see the installation section of the docs for details.

Installation from Source

Obtain the source code from https://github.com/OpenMathLib/OpenBLAS/. Note that the default branch is develop (a master branch is still present, but far out of date).

Build-time parameters can be chosen in Makefile.rule, see there for a short description of each option. Most options can also be given directly on the command line as parameters to your make or cmake invocation.

Dependencies

Building OpenBLAS requires the following to be installed:

• GNU Make or CMake
• A C compiler, e.g. GCC or Clang
• A Fortran compiler (optional, for LAPACK)

In general, using a recent version of the compiler is strongly recommended. If a Fortran compiler is not available, it is possible to compile an older version of the included LAPACK that has been machine-translated to C.

Normal compile

Simply invoking make (or gmake on BSD) will detect the CPU automatically. To set a specific target CPU, use make TARGET=xxx, e.g. make TARGET=NEHALEM. The full target list is in the file TargetList.txt, other build optionss are documented in Makefile.rule and can either be set there (typically by removing the comment character from the respective line), or used on the make command line. Note that when you run make install after building, you need to repeat all command line options you provided to make in the build step, as some settings like the supported maximum number of threads are automatically derived from the build host by default, which might not be what you want. For building with cmake, the usual conventions apply, i.e. create a build directory either underneath the toplevel OpenBLAS source directory or separate from it, and invoke cmake there with the path to the source tree and any build options you plan to set.

For more details, see the Building from source section in the docs.

Cross compile

Set CC and FC to point to the cross toolchains, and if you use make, also set HOSTCC to your host C compiler. The target must be specified explicitly when cross compiling.

Examples:

• On a Linux system, cross-compiling to an older MIPS64 router board:

  make BINARY=64 CC=mipsisa64r6el-linux-gnuabi64-gcc FC=mipsisa64r6el-linux-gnuabi64-gfortran HOSTCC=gcc TARGET=P6600
  

• or to a Windows x64 host:

make CC="i686-w64-mingw32-gcc -Bstatic" FC="i686-w64-mingw32-gfortran -static-libgfortran" TARGET=HASWELL BINARY=32 CROSS=1 NUM_THREADS=20 CONSISTENT_FPCSR=1 HOSTCC=gcc

You can find instructions for other cases both in the "Supported Systems" section below and in the Building from source docs. The .yml scripts included with the sources (which contain the build scripts for the "continuous integration" (CI) build tests automatically run on every proposed change to the sources) may also provide additional hints.

When compiling for a more modern CPU target of the same architecture, e.g. TARGET=SKYLAKEX on a HASWELL host, option CROSS=1 can be used to suppress the automatic invocation of the tests at the end of the build.

Debug version

A debug version can be built using make DEBUG=1.

Compile with MASS support on Power CPU (optional)

The IBM MASS library consists of a set of mathematical functions for C, C++, and Fortran applications that are tuned for optimum performance on POWER architectures. OpenBLAS with MASS requires a 64-bit, little-endian OS on POWER. The library can be installed as shown:

• On Ubuntu:

  wget -q http://public.dhe.ibm.com/software/server/POWER/Linux/xl-compiler/eval/ppc64le/ubuntu/public.gpg -O- | sudo apt-key add -
  echo "deb http://public.dhe.ibm.com/software/server/POWER/Linux/xl-compiler/eval/ppc64le/ubuntu/ trusty main" | sudo tee /etc/apt/sources.list.d/ibm-xl-compiler-eval.list
  sudo apt-get update
  sudo apt-get install libxlmass-devel.8.1.5
  

• On RHEL/CentOS:

  wget http://public.dhe.ibm.com/software/server/POWER/Linux/xl-compiler/eval/ppc64le/rhel7/repodata/repomd.xml.key
  sudo rpm --import repomd.xml.key
  wget http://public.dhe.ibm.com/software/server/POWER/Linux/xl-compiler/eval/ppc64le/rhel7/ibm-xl-compiler-eval.repo
  sudo cp ibm-xl-compiler-eval.repo /etc/yum.repos.d/
  sudo yum install libxlmass-devel.8.1.5
  

After installing the MASS library, compile OpenBLAS with USE_MASS=1. For example, to compile on Power8 with MASS support: make USE_MASS=1 TARGET=POWER8.

Install to a specific directory (optional)

Use PREFIX= when invoking make, for example

make install PREFIX=your_installation_directory

(along with all options you added on the make command line in the preceding build step) The default installation directory is /opt/OpenBLAS.

Supported CPUs and Operating Systems

Please read GotoBLAS_01Readme.txt for older CPU models already supported by the 2010 GotoBLAS.

Additional supported CPUs

x86/x86-64

• Intel Xeon 56xx (Westmere): Used GotoBLAS2 Nehalem codes.
• Intel Sandy Bridge: Optimized Level-3 and Level-2 BLAS with AVX on x86-64.
• Intel Haswell: Optimized Level-3 and Level-2 BLAS with AVX2 and FMA on x86-64.
• Intel Skylake-X: Optimized Level-3 and Level-2 BLAS with AVX512 and FMA on x86-64.
• Intel Cooper Lake: as Skylake-X with improved BFLOAT16 support.
• Intel Sapphire Rapids: as Cooper Lake with improved BFLOAT16 SBGEMM kernel.
• AMD Bobcat: Used GotoBLAS2 Barcelona codes.
• AMD Bulldozer: x86-64 ?GEMM FMA4 kernels. (Thanks to Werner Saar)
• AMD PILEDRIVER: Uses Bulldozer codes with some optimizations.
• AMD STEAMROLLER: Uses Bulldozer codes with some optimizations.
• AMD ZEN: Uses Haswell codes with some optimizations for Zen 2/3, SkylakeX for Zen4, Cooperlake for Zen5

MIPS32

• MIPS 1004K: uses P5600 codes
• MIPS 24K: uses P5600 codes

MIPS64

• ICT Loongson 3A: Optimized Level-3 BLAS and the part of Level-1,2.
• ICT Loongson 3B: Experimental

ARM

• ARMv6: Optimized BLAS for vfpv2 and vfpv3-d16 (e.g. BCM2835, Cortex M0+)
• ARMv7: Optimized BLAS for vfpv3-d32 (e.g. Cortex A8, A9 and A15)

ARM64

• ARMv8: Basic ARMV8 with small caches, optimized Level-3 and Level-2 BLAS
• Cortex-A53: same as ARMV8 (different cpu specifications)
• Cortex-A55: same as ARMV8 (different cpu specifications)
• Cortex A57: Optimized Level-3 and Level-2 functions
• Cortex A72: same as A57 (different cpu specifications)
• Cortex A73: same as A57 (different cpu specifications)
• Cortex A76: same as A57 (different cpu specifications)
• Falkor: same as A57 (different cpu specifications)
• ThunderX: Optimized some Level-1 functions
• ThunderX2T99: Optimized Level-3 BLAS and parts of Levels 1 and 2
• ThunderX3T110
• TSV110: Optimized some Level-3 helper functions
• EMAG 8180: preliminary support based on A57
• Neoverse N1: (AWS Graviton2) preliminary support
• Neoverse V1: (AWS Graviton3) optimized Level-3 BLAS
• Neoverse N2: preliminary support
• Neoverse V2: preliminary support
• Neoverse V3: preliminary support
• Neoverse V3AE: preliminary support
• Apple Vortex: preliminary support based on ThunderX2/3
• Apple VortexM4: preliminary support based on ThunderX2/3, SME kernels for SGEMM,SSYMM,STRMM,SSYRK,SSYR2K
• A64FX: preliminary support, optimized Level-3 BLAS
• ARMV8SVE: any ARMV8 cpu with SVE extensions
• ARMV9SME: any ARMV9 cpu with SVE and SME extensions

PPC/PPC64

• POWER8: Optimized BLAS, only for PPC64LE (Little Endian), only with USE_OPENMP=1

• POWER9: Optimized Level-3 BLAS (real) and some Level-1,2. PPC64LE with OpenMP only.

• POWER10: Optimized Level-3 BLAS including SBGEMM and some Level-1,2.

• AIX: Dynamic architecture with OpenXL and OpenMP.

  make CC=ibm-clang_r FC=xlf_r TARGET=POWER7 BINARY=64 USE_OPENMP=1 INTERFACE64=1 DYNAMIC_ARCH=1 USE_THREAD=1
  

IBM zEnterprise System

• Z13: Optimized Level-3 BLAS and Level-1,2
• Z14: Optimized Level-3 BLAS and (single precision) Level-1,2

RISC-V

• C910V: Optimized Level-3 BLAS (real) and Level-1,2 by RISC-V Vector extension 0.7.1.

  make HOSTCC=gcc TARGET=C910V CC=riscv64-unknown-linux-gnu-gcc FC=riscv64-unknown-linux-gnu-gfortran
  

  (also known to work on C906 as long as you use only single-precision functions - its instruction set support appears to be incomplete in double precision)

• x280: Level-3 BLAS and Level-1,2 are optimized by RISC-V Vector extension 1.0.

  make HOSTCC=gcc TARGET=x280 NUM_THREADS=8 CC=riscv64-unknown-linux-gnu-clang FC=riscv64-unknown-linux-gnu-gfortran
  

(For RVV 1.0 target selection and compiler requirements, see the RISC-V section in docs/install.md.)

• ZVL???B: Level-3 BLAS and Level-1,2 including vectorised kernels targeting generic RISCV cores with vector support with registers of at least the corresponding width; ZVL128B and ZVL256B are available. e.g.:

    make TARGET=RISCV64_ZVL256B CFLAGS="-DTARGET=RISCV64_ZVL256B" \
    BINARY=64 ARCH=riscv64 CC='clang -target riscv64-unknown-linux-gnu' \
    AR=riscv64-unknown-linux-gnu-ar AS=riscv64-unknown-linux-gnu-gcc \
    LD=riscv64-unknown-linux-gnu-gcc FC=riscv64-unknown-linux-gnu-gfortran \
    HOSTCC=gcc HOSTFC=gfortran -j
  

LOONGARCH64

• LA64_GENERIC: Optimized Level-3, Level-2 and Level-1 BLAS with scalar instruction

  make HOSTCC=gcc TARGET=LA64_GENERIC CC=loongarch64-unknown-linux-gnu-gcc FC=loongarch64-unknown-linux-gnu-gfortran USE_SIMPLE_THREADED_LEVEL3=1
  

  The old-style TARGET=LOONGSONGENERIC is still supported

• LA264: Optimized Level-3, Level-2 and Level-1 BLAS with LSX instruction

  make HOSTCC=gcc TARGET=LA264 CC=loongarch64-unknown-linux-gnu-gcc FC=loongarch64-unknown-linux-gnu-gfortran USE_SIMPLE_THREADED_LEVEL3=1
  

  The old-style TARGET=LOONGSON2K1000 is still supported

• LA464: Optimized Level-3, Level-2 and Level-1 BLAS with LASX instruction

  make HOSTCC=gcc TARGET=LA464 CC=loongarch64-unknown-linux-gnu-gcc FC=loongarch64-unknown-linux-gnu-gfortran USE_SIMPLE_THREADED_LEVEL3=1
  

  The old-style TARGET=LOONGSON3R5 is still supported

WASM

Not a cpu target in the strict sense, but portable WebAssembly for browser-based applications and the like. See emscripten.org for the compiler and related information

• WASM128_GENERIC: Optimized SGEMM,DGEMM, DAXPY, SSUM/DSUM, SDOT/DDOT and SROT/DROT

Support for multiple targets in a single library

OpenBLAS can be built for multiple targets with runtime detection of the target cpu by specifying DYNAMIC_ARCH=1 in Makefile.rule, on the gmake command line or as -DDYNAMIC_ARCH=TRUE in cmake.

For x86_64, the list of targets this activates contains Prescott, Core2, Nehalem, Barcelona, Sandybridge, Bulldozer, Piledriver, Steamroller, Excavator, Haswell, Zen, SkylakeX, Cooper Lake, Sapphire Rapids. For cpu generations not included in this list, the corresponding older model is used. If you also specify DYNAMIC_OLDER=1, specific support for Penryn, Dunnington, Opteron, Opteron/SSE3, Bobcat, Atom and Nano is added. Finally there is an option DYNAMIC_LIST that allows to specify an individual list of targets to include instead of the default.

DYNAMIC_ARCH is also supported on x86, where it translates to Katmai, Coppermine, Northwood, Prescott, Banias, Core2, Penryn, Dunnington, Nehalem, Athlon, Opteron, Opteron_SSE3, Barcelona, Bobcat, Atom and Nano.

On ARMV8, it enables support for CortexA53, CortexA57, CortexA72, CortexA73, Falkor, ThunderX, ThunderX2T99, TSV110 as well as generic ARMV8 cpus. If compiler support for SVE is available at build time, support for NeoverseN2, NeoverseV1 as well as generic ArmV8SVE targets is also enabled.

For POWER, the list encompasses POWER6, POWER8 and POWER9. POWER10 is additionally available if a sufficiently recent compiler is used for the build.

on ZARCH it comprises Z13 and Z14 as well as generic zarch support.

On riscv64, DYNAMIC_ARCH enables support for riscv64_zvl128b and riscv64_zvl256b in addition to generic riscv64 support. A compiler that supports RVV 1.0 is required to build OpenBLAS for riscv64 when DYNAMIC_ARCH is enabled.

On LoongArch64, it comprises LA264 and LA464 as well as generic LoongArch64 support.

The TARGET option can - and usually should - be used in conjunction with DYNAMIC_ARCH=1 to specify which cpu model should be assumed for all the common code in the library, usually you will want to set this to the oldest model you expect to encounter. Failure to specify this may lead to advanced instructions being used by the compiler, just because the build host happens to support them. This is most likely to happen when aggressive optimization options are in effect, and the resulting library may then crash with an illegal instruction error on weaker hardware, before it even reaches the BLAS routines specifically included for that cpu.

Please note that it is not possible to combine support for different architectures, so no combined 32 and 64 bit or x86_64 and arm64 in the same library.

Supported OS

• GNU/Linux
• MinGW or Visual Studio (CMake)/Windows: Please read https://github.com/OpenMathLib/OpenBLAS/blob/develop/docs/install.md#visual-studio--native-windows-abi.
• Darwin/macOS/OSX/iOS: Already supported on PPC and x86 by the original GotoBLAS, now also on ARM64 but we are not OSX/iOS experts.
• FreeBSD: Supported by the community. Basic test coverage is provided by GitHub Actions.
• OpenBSD: Supported by the community. We don't actively test the library on this OS.
• NetBSD: Supported by the community. We don't actively test the library on this OS.
• DragonFly BSD: Supported by the community. We don't actively test the library on this OS.
• Android: Supported by the community. Please read https://github.com/OpenMathLib/OpenBLAS/blob/develop/docs/install.md#android.
• AIX: Supported on PPC up to POWER10 but testing is increasingly problematic due to lack of publicly available systems
• Haiku: Supported by the community. We don't actively test the library on this OS.
• SunOS: Supported by the community. We don't actively test the library on this OS.
• Cortex-M: Supported by the community. Please read https://github.com/OpenMathLib/OpenBLAS/blob/develop/docs/install.md#cortex-m.

Usage

Statically link with libopenblas.a or dynamically link with -lopenblas if OpenBLAS was compiled as a shared library.

Considerations for using the library from Java

The default stack size of only 1MB may be too small, especially if you built OpenBLAS to support larger matrix sizes than provided for by the default settings. Use the -Xss option to request a larger stack size if you encounter problems.

When a Windows build of OpenBLAS was created using the MINGW gfortran (for the LAPACK parts), the java application may hang on startup due to a deadlock between the gfortran runtime library initialization and any pipes created by a Win11/SBT/Play Framework environment. Use -Djdk.console=jdk.internal.le to work around this.

Setting the number of threads using environment variables

Environment variables are used to specify a maximum number of threads. For example,

export OPENBLAS_NUM_THREADS=4
export GOTO_NUM_THREADS=4
export OMP_NUM_THREADS=4

The priorities are OPENBLAS_NUM_THREADS > GOTO_NUM_THREADS > OMP_NUM_THREADS.

If you compile this library with USE_OPENMP=1, you should set the OMP_NUM_THREADS environment variable; OpenBLAS ignores OPENBLAS_NUM_THREADS and GOTO_NUM_THREADS when compiled with USE_OPENMP=1.

Setting the number of threads at runtime

We provide the following functions to control the number of threads at runtime:

void goto_set_num_threads(int num_threads);
void openblas_set_num_threads(int num_threads);

Note that these are only used once at library initialization, and are not available for fine-tuning thread numbers in individual BLAS calls. If you compile this library with USE_OPENMP=1, you should use the above functions too.

Reporting bugs

Please submit an issue in https://github.com/OpenMathLib/OpenBLAS/issues.

Contact

• Use github discussions: https://github.com/OpenMathLib/OpenBLAS/discussions

• OpenBLAS users mailing list: https://groups.google.com/forum/#!forum/openblas-users
• OpenBLAS developers mailing list: https://groups.google.com/forum/#!forum/openblas-dev

Change log

Please see Changelog.txt.

Troubleshooting

• Please read the FAQ section of the docs first.
• Please use GCC version 4.6 and above to compile Sandy Bridge AVX kernels on Linux/MinGW/BSD.
• Please use Clang version 3.1 and above to compile the library on Sandy Bridge microarchitecture. Clang 3.0 will generate the wrong AVX binary code.
• Please use GCC version 6 or LLVM version 6 and above to compile Skylake/CooperLake AVX512 kernels
• Please use LLVM version 18 and above (version 19 and above on Windows) if you plan to use its new flang compiler for Fortran
• Please use GCC version 11 and above to compile OpenBLAS on the POWER architecture
• The number of CPUs/cores should be less than or equal to 256. On Linux x86_64 (amd64), there is experimental support for up to 1024 CPUs/cores and 128 numa nodes if you build the library with BIGNUMA=1.
• OpenBLAS does not set processor affinity by default. On Linux, you can enable processor affinity by commenting out the line NO_AFFINITY=1 in Makefile.rule. However, note that this may cause a conflict with R parallel.
• On Loongson 3A, make test may fail with a pthread_create error (EAGAIN). However, it will be okay when you run the same test case on the shell.

Contributing

1. Check for open issues or open a fresh issue to start a discussion around a feature idea or a bug.
2. Fork the OpenBLAS repository to start making your changes.
3. Write a test which shows that the bug was fixed or that the feature works as expected.
4. Send a pull request. Make sure to add yourself to CONTRIBUTORS.md.

Donation

Please see the donations section in the docs.