MIN_IOS_VERSION as necessary for your installation. E.g., change the version number
to the minimum iOS version you want to target and execute this file to build the library.
+For this target you will need the cross-compiler toolchain package by Huawei, which contains solutions for both Windows and Linux. Only the Linux-based +toolchain has been tested so far, but the following instructions may apply similarly to Windows:
+Download https://repo.huaweicloud.com/harmonyos/os/4.1.1-Release/ohos-sdk-windows_linux-public.tar.gz (or whatever newer version may be available in the future). Use tar xvf ohos-sdk-windows_linux_public.tar.gz to unpack it somewhere on your system. This will create a folder named "ohos-sdk" with subfolders "linux" and "windows". In the linux one you will find a ZIP archive named "native-linux-x64-4.1.7.8-Release.zip" - you need to unzip this where you want to +install the cross-compiler, for example in /opt/ohos-sdk.
+In the directory where you unpacked OpenBLAS, create a build directory for cmake, and change into it : +
mkdir build
+cd build
+cmake that came with the SDK, and specify the location of its toolchain file as a cmake option. Also set the build target for OpenBLAS to ARMV8 and specify NOFORTRAN=1 (at least as of version 4.1.1, the SDK contains no Fortran compiler):
+/opt/ohos-sdk/linux/native/build-tools/cmake/bin/cmake -DCMAKE_TOOLCHAIN_FILE=/opt/ohos-sdk/linux/native/build/cmake/ohos.toolchain.cmake \
+ -DOHOS_ARCH="arm64-v8a" -DTARGET=ARMV8 -DNOFORTRAN=1 ..
+/opt/ohos-sdk/linux/native/build-tools/cmake/bin/cmake --build .
+For MIPS targets you will need latest toolchains:
malloc for embedded platforms.
- December 26, 2024
+ January 3, 2025
OpenBLAS is an optimized Basic Linear Algebra Subprograms (BLAS) library based on GotoBLAS2 1.13 BSD version.
OpenBLAS implements low-level routines for performing linear algebra operations such as vector addition, scalar multiplication, dot products, linear combinations, and matrix multiplication. OpenBLAS makes these routines available on multiple platforms, covering server, desktop and mobile operating systems, as well as different architectures including x86, ARM, MIPS, PPC, RISC-V, and zarch.
The old GotoBLAS documentation can be found on GitHub.
"},{"location":"#license","title":"License","text":"OpenBLAS is licensed under the 3-clause BSD license. The full license can be found on GitHub.
"},{"location":"about/","title":"About","text":""},{"location":"about/#mailing-list","title":"Mailing list","text":"We have a GitHub discussions forum to discuss usage and development of OpenBLAS. We also have a Google group for users and a Google group for development of OpenBLAS.
"},{"location":"about/#acknowledgements","title":"Acknowledgements","text":"This work was or is partially supported by the following grants, contracts and institutions:
Over the course of OpenBLAS development, a number of donations were received. You can read OpenBLAS's statement of receipts and disbursement and cash balance in this Google doc (covers 2013-2016). A list of backers is available in BACKERS.md in the main repo.
"},{"location":"about/#donations","title":"Donations","text":"We welcome hardware donations, including the latest CPUs and motherboards.
"},{"location":"about/#open-source-users-of-openblas","title":"Open source users of OpenBLAS","text":"Prominent open source users of OpenBLAS include:
OpenBLAS is packaged in most major Linux distros, as well as general and numerical computing-focused packaging ecosystems like Nix, Homebrew, Spack and conda-forge.
OpenBLAS is used directly by libraries written in C, C++ and Fortran (and probably other languages), and directly by end users in those languages.
"},{"location":"about/#publications","title":"Publications","text":""},{"location":"about/#2013","title":"2013","text":"This page describes the Make-based build, which is the default/authoritative build method. Note that the OpenBLAS repository also supports building with CMake (not described here) - that generally works and is tested, however there may be small differences between the Make and CMake builds.
Warning
This page is made by someone who is not the developer and should not be considered as an official documentation of the build system. For getting the full picture, it is best to read the Makefiles and understand them yourself.
"},{"location":"build_system/#makefile-dep-graph","title":"Makefile dep graph","text":"Makefile \n| \n|----- Makefile.system # !!! this is included by many of the Makefiles in the subdirectories !!!\n| |\n| |===== Makefile.prebuild # This is triggered (not included) once by Makefile.system \n| | | # and runs before any of the actual library code is built.\n| | | # (builds and runs the \"getarch\" tool for cpu identification,\n| | | # runs the compiler detection scripts c_check and f_check) \n| | |\n| | ----- (Makefile.conf) [ either this or Makefile_kernel.conf is generated ] \n| | | { Makefile.system#L243 }\n| | ----- (Makefile_kernel.conf) [ temporary Makefile.conf during DYNAMIC_ARCH builds ]\n| |\n| |----- Makefile.rule # defaults for build options that can be given on the make command line\n| |\n| |----- Makefile.$(ARCH) # architecture-specific compiler options and OpenBLAS buffer size values\n|\n|~~~~~ exports/\n|\n|~~~~~ test/\n|\n|~~~~~ utest/ \n|\n|~~~~~ ctest/\n|\n|~~~~~ cpp_thread_test/\n|\n|~~~~~ kernel/\n|\n|~~~~~ ${SUBDIRS}\n|\n|~~~~~ ${BLASDIRS}\n|\n|~~~~~ ${NETLIB_LAPACK_DIR}{,/timing,/testing/{EIG,LIN}}\n|\n|~~~~~ relapack/\nMost of the tunable variables are found in Makefile.rule, along with their detailed descriptions. Most of the variables are detected automatically in Makefile.prebuild, if they are not set in the environment.
"},{"location":"build_system/#cpu-related","title":"CPU related","text":"ARCH - Target architecture (eg. x86_64)\nTARGET - Target CPU architecture, in case of DYNAMIC_ARCH=1 means library will not be usable on less capable CPUs\nTARGET_CORE - TARGET_CORE will override TARGET internally during each cpu-specific cycle of the build for DYNAMIC_ARCH\nDYNAMIC_ARCH - For building library for multiple TARGETs (does not lose any optimizations, but increases library size)\nDYNAMIC_LIST - optional user-provided subset of the DYNAMIC_CORE list in Makefile.system\nCC - TARGET C compiler used for compilation (can be cross-toolchains)\nFC - TARGET Fortran compiler used for compilation (can be cross-toolchains, set NOFORTRAN=1 if used cross-toolchain has no fortran compiler)\nAR, AS, LD, RANLIB - TARGET toolchain helpers used for compilation (can be cross-toolchains)\n\nHOSTCC - compiler of build machine, needed to create proper config files for target architecture\nHOST_CFLAGS - flags for build machine compiler\nBINARY - 32/64 bit library\n\nBUILD_SHARED - Create shared library\nBUILD_STATIC - Create static library\n\nQUAD_PRECISION - enable support for IEEE quad precision [ largely unimplemented leftover from GotoBLAS, do not use ]\nEXPRECISION - Obsolete option to use float80 of SSE on BSD-like systems\nINTERFACE64 - Build with 64bit integer representations to support large array index values [ incompatible with standard API ]\n\nBUILD_SINGLE - build the single-precision real functions of BLAS [and optionally LAPACK] \nBUILD_DOUBLE - build the double-precision real functions\nBUILD_COMPLEX - build the single-precision complex functions\nBUILD_COMPLEX16 - build the double-precision complex functions\n(all four types are included in the build by default when none was specifically selected)\n\nBUILD_BFLOAT16 - build the \"half precision brainfloat\" real functions \n\nUSE_THREAD - Use a multithreading backend (default to pthread)\nUSE_LOCKING - implement locking for thread safety even when USE_THREAD is not set (so that the singlethreaded library can\n safely be called from multithreaded programs)\nUSE_OPENMP - Use OpenMP as multithreading backend\nNUM_THREADS - define this to the maximum number of parallel threads you expect to need (defaults to the number of cores in the build cpu)\nNUM_PARALLEL - define this to the number of OpenMP instances that your code may use for parallel calls into OpenBLAS (default 1,see below)\nOpenBLAS uses a fixed set of memory buffers internally, used for communicating and compiling partial results from individual threads. For efficiency, the management array structure for these buffers is sized at build time - this makes it necessary to know in advance how many threads need to be supported on the target system(s).
With OpenMP, there is an additional level of complexity as there may be calls originating from a parallel region in the calling program. If OpenBLAS gets called from a single parallel region, it runs single-threaded automatically to avoid overloading the system by fanning out its own set of threads. In the case that an OpenMP program makes multiple calls from independent regions or instances in parallel, this default serialization is not sufficient as the additional caller(s) would compete for the original set of buffers already in use by the first call. So if multiple OpenMP runtimes call into OpenBLAS at the same time, then only one of them will be able to make progress while all the rest of them spin-wait for the one available buffer. Setting NUM_PARALLEL to the upper bound on the number of OpenMP runtimes that you can have in a process ensures that there are a sufficient number of buffer sets available.
OpenBLAS/ \n\u251c\u2500\u2500 benchmark Benchmark codes for BLAS\n\u251c\u2500\u2500 cmake CMakefiles\n\u251c\u2500\u2500 ctest Test codes for CBLAS interfaces\n\u251c\u2500\u2500 driver Implemented in C\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 level2\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 level3\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 mapper\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 others Memory management, threading, etc\n\u251c\u2500\u2500 exports Generate shared library\n\u251c\u2500\u2500 interface Implement BLAS and CBLAS interfaces (calling driver or kernel)\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 lapack\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 netlib\n\u251c\u2500\u2500 kernel Optimized assembly kernels for CPU architectures\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 alpha Original GotoBLAS kernels for DEC Alpha\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 arm ARMV5,V6,V7 kernels (including generic C codes used by other architectures)\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 arm64 ARMV8\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 generic General kernel codes written in plain C, parts used by many architectures.\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 ia64 Original GotoBLAS kernels for Intel Itanium\n\u2502 \u251c\u2500\u2500 mips\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 mips64\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 power\n| \u251c\u2500\u2500 riscv64\n| \u251c\u2500\u2500 simd Common code for Universal Intrinsics, used by some x86_64 and arm64 kernels\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 sparc\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 x86\n\u2502 \u251c\u2500\u2500 x86_64\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 zarch \n\u251c\u2500\u2500 lapack Optimized LAPACK codes (replacing those in regular LAPACK)\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 getf2\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 getrf\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 getrs\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 laswp\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 lauu2\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 lauum\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 potf2\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 potrf\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 trti2\n\u2502 \u251c\u2500\u2500 trtri\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 trtrs\n\u251c\u2500\u2500 lapack-netlib LAPACK codes from netlib reference implementation\n\u251c\u2500\u2500 reference BLAS Fortran reference implementation (unused)\n\u251c\u2500\u2500 relapack Elmar Peise's recursive LAPACK (implemented on top of regular LAPACK)\n\u251c\u2500\u2500 test Test codes for BLAS\n\u2514\u2500\u2500 utest Regression test\nA call tree for dgemm looks as follows:
interface/gemm.c\n \u2502\ndriver/level3/level3.c\n \u2502\ngemm assembly kernels at kernel/\nTo find the kernel currently used for a particular supported CPU, please check the corresponding kernel/$(ARCH)/KERNEL.$(CPU) file.
Here is an example for kernel/x86_64/KERNEL.HASWELL:
...\nDTRMMKERNEL = dtrmm_kernel_4x8_haswell.c\nDGEMMKERNEL = dgemm_kernel_4x8_haswell.S\n...\nKERNEL.HASWELL, OpenBLAS Haswell dgemm kernel file is dgemm_kernel_4x8_haswell.S."},{"location":"developers/#optimizing-gemm-for-a-given-hardware","title":"Optimizing GEMM for a given hardware","text":"Read the Goto paper to understand the algorithm
Goto, Kazushige; van de Geijn, Robert A. (2008). \"Anatomy of High-Performance Matrix Multiplication\". ACM Transactions on Mathematical Software 34 (3): Article 12
(The above link is available only to ACM members, but this and many related papers is also available on the pages of van de Geijn's FLAME project)
The driver/level3/level3.c is the implementation of Goto's algorithm. Meanwhile, you can look at kernel/generic/gemmkernel_2x2.c, which is a naive 2x2 register blocking gemm kernel in C. Then:
Mc, Kc, and Nc)Note that not all of the CPU-specific parameters in param.h are actively used in algorithms. DNUMOPT only appears as a scale factor in profiling output of the level3 syrk interface code, while its counterpart SNUMOPT (aliased as NUMOPT in common.h) is not used anywhere at all.
SYMV_P is only used in the generic kernels for the symv and chemv/zhemv functions - at least some of those are usually overridden by CPU-specific implementations, so if you start by cloning the existing implementation for a related CPU you need to check its KERNEL file to see if tuning SYMV_P would have any effect at all.
GEMV_UNROLL is only used by some older x86-64 kernels, so not all sections in param.h define it. Similarly, not all of the CPU parameters like L2 or L3 cache sizes are necessarily used in current kernels for a given model - by all indications the CPU identification code was imported from some other project originally.
We use tests for Netlib BLAS, CBLAS, and LAPACK. In addition, we use OpenBLAS-specific regression tests. They can be run with Make:
make -C test for BLAS testsmake -C ctest for CBLAS testsmake -C utest for OpenBLAS regression testsmake lapack-test for LAPACK testsWe also use the BLAS-Tester tests for regression testing. It is basically the ATLAS test suite adapted for building with OpenBLAS.
The project makes use of several Continuous Integration (CI) services conveniently interfaced with GitHub to automatically run tests on a number of platforms and build configurations.
Also note that the test suites included with \"numerically heavy\" projects like Julia, NumPy, SciPy, Octave or QuantumEspresso can be used for regression testing, when those projects are built such that they use OpenBLAS.
"},{"location":"developers/#benchmarking","title":"Benchmarking","text":"A number of benchmarking methods are used by OpenBLAS:
benchmark folder. They can be run locally through the Makefile in that directory. And the benchmark/scripts subdirectory contains similar benchmarks that use OpenBLAS via NumPy, SciPy, Octave and R.Benchmarking code for BLAS libraries, and specific performance analysis results, can be found in a number of places. For example:
Especially relevant for x86-64, a new CPU model may be a \"refresh\" (die shrink and/or different number of cores) within an existing model family without significant changes to its instruction set (e.g., Intel Skylake and Kaby Lake still are fundamentally the same architecture as Haswell, low end Goldmont etc. are Nehalem). In this case, compilation with the appropriate older TARGET will already lead to a satisfactory build.
To achieve autodetection of the new model, its CPUID (or an equivalent identifier) needs to be added in the cpuid_<architecture>.c relevant for its general architecture, with the returned name for the new type set appropriately. For x86, which has the most complex cpuid file, there are two functions that need to be edited: get_cpuname() to return, e.g., CPUTYPE_HASWELL and get_corename() for the (broader) core family returning, e.g., CORE_HASWELL.1
For architectures where DYNAMIC_ARCH builds are supported, a similar but simpler code section for the corresponding runtime detection of the CPU exists in driver/others/dynamic.c (for x86), and driver/others/dynamic_<arch>.c for other architectures. Note that for x86 the CPUID is compared after splitting it into its family, extended family, model and extended model parts, so the single decimal number returned by Linux in /proc/cpuinfo for the model has to be converted back to hexadecimal before splitting into its constituent digits. For example, 142 == 8E translates to extended model 8, model 14.
Usually it will be possible to start from an existing model, clone its KERNEL configuration file to the new name to use for this TARGET and eventually replace individual kernels with versions better suited for peculiarities of the new CPU model. In addition, it is necessary to add (or clone at first) the corresponding section of GEMM_UNROLL parameters in the top-level param.h, and possibly to add definitions such as USE_TRMM (governing whether TRMM functions use the respective GEMM kernel or a separate source file) to the Makefiles (and CMakeLists.txt) in the kernel directory. The new CPU name needs to be added to TargetList.txt, and the CPU auto-detection code used by the getarch helper program - contained in the cpuid_<architecture>.c file amended to include the CPUID (or equivalent) information processing required (see preceding section).
This endeavour is best started by cloning the entire support structure for 32-bit ARM, and within that the ARMv5 CPU in particular, as this is implemented through plain C kernels only. An example providing a convenient \"shopping list\" can be seen in pull request #1526.
This information ends up in the Makefile.conf and config.h files generated by getarch. Failure to set either will typically lead to a missing definition of the GEMM_UNROLL parameters later in the build, as getarch_2nd will be unable to find a matching parameter section in param.h.\u00a0\u21a9
Note
This document contains recommendations only - packagers and other redistributors are in charge of how OpenBLAS is built and distributed in their systems, and may have good reasons to deviate from the guidance given on this page. These recommendations are aimed at general packaging systems, with a user base that typically is large, open source (or freely available at least), and doesn't behave uniformly or that the packager is directly connected with.*
OpenBLAS has a large number of build-time options which can be used to change how it behaves at runtime, how artifacts or symbols are named, etc. Variation in build configuration can be necessary to acheive a given end goal within a distribution or as an end user. However, such variation can also make it more difficult to build on top of OpenBLAS and ship code or other packages in a way that works across many different distros. Here we provide guidance about the most important build options, what effects they may have when changed, and which ones to default to.
The Make and CMake build systems provide equivalent options and yield more or less the same artifacts, but not exactly (the CMake builds are still experimental). You can choose either one and the options will function in the same way, however the CMake outputs may require some renaming. To review available build options, see Makefile.rule or CMakeLists.txt in the root of the repository.
Build options typically fall into two categories: (a) options that affect the user interface, such as library and symbol names or APIs that are made available, and (b) options that affect performance and runtime behavior, such as threading behavior or CPU architecture-specific code paths. The user interface options are more important to keep aligned between distributions, while for the performance-related options there are typically more reasons to make choices that deviate from the defaults.
Here are recommendations for user interface related packaging choices where it is not likely to be a good idea to deviate (typically these are the default settings):
.pc) and CMake .cmake) dependency detection files. These files are used by build systems when users want to link against OpenBLAS, and there is no benefit of leaving them out.The LP64 (32-bit integer) interface is the default build, and has well-established C and Fortran APIs as determined by the reference (Netlib) BLAS and LAPACK libraries. The ILP64 (64-bit integer) interface however does not have a standard API: symbol names and shared/static library names can be produced in multiple ways, and this tends to make it difficult to use. As of today there is an agreed-upon way of choosing names for OpenBLAS between a number of key users/redistributors, which is the closest thing to a standard that there is now. However, there is an ongoing standardization effort in the reference BLAS and LAPACK libraries, which differs from the current OpenBLAS agreed-upon convention. In this section we'll aim to explain both.
Those two methods are fairly similar, and have a key thing in common: using a symbol suffix. This is good practice; it is recommended that if you distribute an ILP64 build, to have it use a symbol suffix containing 64 in the name. This avoids potential symbol clashes when different packages which depend on OpenBLAS load both an LP64 and an ILP64 library into memory at the same time.
This convention comprises the shared library name and the symbol suffix in the shared library. The symbol suffix to use is 64_, implying that the library name will be libopenblas64_.so and the symbols in that library end in 64_. The central issue where this was discussed is openblas#646, and adopters include Fedora, Julia, NumPy and SciPy - SuiteSparse already used it as well.
To build shared and static libraries with the currently recommended ILP64 conventions with Make:
$ make INTERFACE64=1 SYMBOLSUFFIX=64_\nThis will produce libraries named libopenblas64_.so|a, a pkg-config file named openblas64.pc, and CMake and header files.
Installing locally and inspecting the output will show a few more details:
$ make install PREFIX=$PWD/../openblas/make64 INTERFACE64=1 SYMBOLSUFFIX=64_\n$ tree . # output slightly edited down\n.\n\u251c\u2500\u2500 include\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 cblas.h\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 f77blas.h\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 lapacke_config.h\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 lapacke.h\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 lapacke_mangling.h\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 lapacke_utils.h\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 lapack.h\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 openblas_config.h\n\u2514\u2500\u2500 lib\n \u251c\u2500\u2500 cmake\n \u2502\u00a0\u00a0 \u2514\u2500\u2500 openblas\n \u2502\u00a0\u00a0 \u251c\u2500\u2500 OpenBLASConfig.cmake\n \u2502\u00a0\u00a0 \u2514\u2500\u2500 OpenBLASConfigVersion.cmake\n \u251c\u2500\u2500 libopenblas64_.a\n \u251c\u2500\u2500 libopenblas64_.so\n \u2514\u2500\u2500 pkgconfig\n \u2514\u2500\u2500 openblas64.pc\nA key point are the symbol names. These will equal the LP64 symbol names, then (for Fortran only) the compiler mangling, and then the 64_ symbol suffix. Hence to obtain the final symbol names, we need to take into account which Fortran compiler we are using. For the most common cases (e.g., gfortran, Intel Fortran, or Flang), that means appending a single underscore. In that case, the result is:
dgemm dgemm_64_ dgemm_64(...) dgemm_64_(...) cblas_dgemm cblas_dgemm64_ n/a cblas_dgemm64_(...) It is quite useful to have these symbol names be as uniform as possible across different packaging systems.
The equivalent build options with CMake are:
$ mkdir build && cd build\n$ cmake .. -DINTERFACE64=1 -DSYMBOLSUFFIX=64_ -DBUILD_SHARED_LIBS=ON -DBUILD_STATIC_LIBS=ON\n$ cmake --build . -j\nNote that the result is not 100% identical to the Make result. For example, the library name ends in _64 rather than 64_ - it is recommended to rename them to match the Make library names (also update the libsuffix entry in openblas64.pc to match that rename).
$ cmake --install . --prefix $PWD/../../openblas/cmake64\n$ tree .\n.\n\u251c\u2500\u2500 include\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 openblas64\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 cblas.h\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 f77blas.h\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 lapacke_config.h\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 lapacke_example_aux.h\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 lapacke.h\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 lapacke_mangling.h\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 lapacke_utils.h\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 lapack.h\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 openblas64\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u2514\u2500\u2500 lapacke_mangling.h\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 openblas_config.h\n\u2514\u2500\u2500 lib\n \u251c\u2500\u2500 cmake\n \u2502\u00a0\u00a0 \u2514\u2500\u2500 OpenBLAS64\n \u2502\u00a0\u00a0 \u251c\u2500\u2500 OpenBLAS64Config.cmake\n \u2502\u00a0\u00a0 \u251c\u2500\u2500 OpenBLAS64ConfigVersion.cmake\n \u2502\u00a0\u00a0 \u251c\u2500\u2500 OpenBLAS64Targets.cmake\n \u2502\u00a0\u00a0 \u2514\u2500\u2500 OpenBLAS64Targets-noconfig.cmake\n \u251c\u2500\u2500 libopenblas_64.a\n \u251c\u2500\u2500 libopenblas_64.so -> libopenblas_64.so.0\n \u2514\u2500\u2500 pkgconfig\n \u2514\u2500\u2500 openblas64.pc\nWhile the 64_ convention above got some adoption, it's slightly hacky and is implemented through the use of objcopy. An effort is ongoing for a more broadly adopted convention in the reference BLAS and LAPACK libraries, using (a) the _64 suffix, and (b) applying that suffix before rather than after Fortran compiler mangling. The central issue for this is lapack#666.
For the most common cases of compiler mangling (a single _ appended), the end result will be:
dgemm dgemm_64_ dgemm_64(...) dgemm_64_(...) cblas_dgemm cblas_dgemm_64 n/a cblas_dgemm_64(...) For other compiler mangling schemes, replace the trailing _ by the scheme in use.
The shared library name for this _64 convention should be libopenblas_64.so.
Note: it is not yet possible to produce an OpenBLAS build which employs this convention! Once reference BLAS and LAPACK with support for _64 have been released, a future OpenBLAS release will support it. For now, please use the older 64_ scheme and avoid using the name libopenblas_64.so; it should be considered reserved for future use of the _64 standard as prescribed by reference BLAS/LAPACK.
For these options there are multiple reasonable or common choices.
"},{"location":"distributing/#threading-related-options","title":"Threading related options","text":"OpenBLAS can be built as a multi-threaded or single-threaded library, with the default being multi-threaded. It's expected that the default libopenblas library is multi-threaded; if you'd like to also distribute single-threaded builds, consider naming them libopenblas_sequential.
OpenBLAS can be built with pthreads or OpenMP as the threading model, with the default being pthreads. Both options are commonly used, and the choice here should not influence the shared library name. The choice will be captured by the .pc file. E.g.,:
$ pkg-config --libs openblas\n-fopenmp -lopenblas\n\n$ cat openblas.pc\n...\nopenblas_config= ... USE_OPENMP=0 MAX_THREADS=24\nThe maximum number of threads users will be able to use is determined at build time by the NUM_THREADS build option. It defaults to 24, and there's a wide range of values that are reasonable to use (up to 256). 64 is a typical choice here; there is a memory footprint penalty that is linear in NUM_THREADS. Please see Makefile.rule for more details.
OpenBLAS contains a lot of CPU architecture-specific optimizations, hence when distributing to a user base with a variety of hardware, it is recommended to enable CPU architecture runtime detection. This will dynamically select optimized kernels for individual APIs. To do this, use the DYNAMIC_ARCH=1 build option. This is usually done on all common CPU families, except when there are known issues.
In case the CPU architecture is known (e.g. you're building binaries for macOS M1 users), it is possible to specify the target architecture directly with the TARGET= build option.
DYNAMIC_ARCH and TARGET are covered in more detail in the main README.md in this repository.
OpenBLAS is likely to be distributed in one of these distribution models:
The guidance on this page is most important for models (1) and (2). These links to build recipes for a representative selection of packaging systems may be helpful as a reference:
All major distributions do include LAPACK as of mid 2023 as far as we know. Older versions of Arch Linux did not, and that was known to cause problems.\u00a0\u21a9
OpenBLAS for the most part contains implementations of the reference (Netlib) BLAS, CBLAS, LAPACK and LAPACKE interfaces. A few OpenBLAS-specific functions are also provided however, which mostly can be seen as \"BLAS extensions\". This page documents those non-standard APIs.
"},{"location":"extensions/#blas-like-extensions","title":"BLAS-like extensions","text":"Routine Data Types Description ?axpby s,d,c,z likeaxpy with a multiplier for y ?gemm3m c,z gemm3m ?imatcopy s,d,c,z in-place transposition/copying ?omatcopy s,d,c,z out-of-place transposition/copying ?geadd s,d,c,z ATLAS-like matrix add B = α*A+β*B ?gemmt s,d,c,z gemm but only a triangular part updated"},{"location":"extensions/#bfloat16-functionality","title":"bfloat16 functionality","text":"BLAS-like and conversion functions for bfloat16 (available when OpenBLAS was compiled with BUILD_BFLOAT16=1):
void cblas_sbstobf16 converts a float array to an array of bfloat16 values by roundingvoid cblas_sbdtobf16 converts a double array to an array of bfloat16 values by roundingvoid cblas_sbf16tos converts a bfloat16 array to an array of floatsvoid cblas_dbf16tod converts a bfloat16 array to an array of doublesfloat cblas_sbdot computes the dot product of two bfloat16 arraysvoid cblas_sbgemv performs the matrix-vector operations of GEMV with the input matrix and X vector as bfloat16void cblas_sbgemm performs the matrix-matrix operations of GEMM with both input arrays containing bfloat16openblas_get_num_threadsopenblas_set_num_threadsint openblas_get_num_procs(void) returns the number of processors available on the system (may include \"hyperthreading cores\")int openblas_get_parallel(void) returns 0 for sequential use, 1 for platform-based threading and 2 for OpenMP-based threadingchar * openblas_get_config() returns the options OpenBLAS was built with, something like NO_LAPACKE DYNAMIC_ARCH NO_AFFINITY Haswellint openblas_set_affinity(int thread_index, size_t cpusetsize, cpu_set_t *cpuset) sets the CPU affinity mask of the given thread to the provided cpuset. Only available on Linux, with semantics identical to pthread_setaffinity_np.BLAS stands for Basic Linear Algebra Subprograms. BLAS provides standard interfaces for linear algebra, including BLAS1 (vector-vector operations), BLAS2 (matrix-vector operations), and BLAS3 (matrix-matrix operations). In general, BLAS is the computational kernel (\"the bottom of the food chain\") in linear algebra or scientific applications. Thus, if BLAS implementation is highly optimized, the whole application can get substantial benefit.
"},{"location":"faq/#what-functions-are-there-and-how-can-i-call-them-from-my-c-code","title":"What functions are there and how can I call them from my C code?","text":"As BLAS is a standardized interface, you can refer to the documentation of its reference implementation at netlib.org. Calls from C go through its CBLAS interface, so your code will need to include the provided cblas.h in addition to linking with -lopenblas. A single-precision matrix multiplication will look like
#include <cblas.h>\n...\ncblas_sgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, M, N, K, 1.0, A, K, B, N, 0.0, result, N);\nOpenBLAS is an open source BLAS library forked from the GotoBLAS2-1.13 BSD version. Since Mr. Kazushige Goto left TACC, GotoBLAS is no longer being maintained. Thus, we created this project to continue developing OpenBLAS/GotoBLAS.
"},{"location":"faq/#whats-the-difference-between-openblas-and-gotoblas","title":"What's the difference between OpenBLAS and GotoBLAS?","text":"In OpenBLAS 0.2.0, we optimized level 3 BLAS on the Intel Sandy Bridge 64-bit OS. We obtained a performance comparable with that Intel MKL.
We optimized level 3 BLAS performance on the ICT Loongson-3A CPU. It outperformed GotoBLAS by 135% in a single thread and 120% in 4 threads.
We fixed some GotoBLAS bugs including a SEGFAULT bug on the new Linux kernel, MingW32/64 bugs, and a ztrmm computing error bug on Intel Nehalem.
We also added some minor features, e.g. supporting \"make install\", compiling without LAPACK and upgrading the LAPACK version to 3.4.2.
You can find the full list of modifications in Changelog.txt.
"},{"location":"faq/#where-do-parameters-gemm_p-gemm_q-gemm_r-come-from","title":"Where do parameters GEMM_P, GEMM_Q, GEMM_R come from?","text":"The detailed explanation is probably in the original publication authored by Kazushige Goto - Goto, Kazushige; van de Geijn, Robert A; Anatomy of high-performance matrix multiplication. ACM Transactions on Mathematical Software (TOMS). Volume 34 Issue 3, May 2008 While this article is paywalled and too old for preprints to be available on arxiv.org, more recent publications like https://arxiv.org/pdf/1609.00076 contain at least a brief description of the algorithm. In practice, the values are derived by experimentation to yield the block sizes that give the highest performance. A general rule of thumb for selecting a starting point seems to be that PxQ is about half the size of L2 cache.
"},{"location":"faq/#how-can-i-report-a-bug","title":"How can I report a bug?","text":"Please file an issue at this issue page or send mail to the OpenBLAS mailing list.
Please provide the following information: CPU, OS, compiler, OpenBLAS version and any compiling flags you used (Makefile.rule). In addition, please describe how to reproduce this bug.
"},{"location":"faq/#how-to-reference-openblas","title":"How to reference OpenBLAS.","text":"You can reference our papers in this page. Alternatively, you can cite the OpenBLAS homepage http://www.openblas.net.
"},{"location":"faq/#how-can-i-use-openblas-in-multi-threaded-applications","title":"How can I use OpenBLAS in multi-threaded applications?","text":"If your application is already multi-threaded, it will conflict with OpenBLAS multi-threading. Thus, you must set OpenBLAS to use single thread as following.
If the application is parallelized by OpenMP, please build OpenBLAS with USE_OPENMP=1
With the increased availability of fast multicore hardware it has unfortunately become clear that the thread management provided by OpenMP is not sufficient to prevent race conditions when OpenBLAS was built single-threaded by USE_THREAD=0 and there are concurrent calls from multiple threads to OpenBLAS functions. In this case, it is vital to also specify USE_LOCKING=1 (introduced with OpenBLAS 0.3.7).
"},{"location":"faq/#does-openblas-support-sparse-matrices-andor-vectors","title":"Does OpenBLAS support sparse matrices and/or vectors ?","text":"OpenBLAS implements only the standard (dense) BLAS and LAPACK functions with a select few extensions popularized by Intel's MKL. Some cases can probably be made to work using e.g. GEMV or AXPBY, in general using a dedicated package like SuiteSparse (which can make use of OpenBLAS or equivalent for standard operations) is recommended.
"},{"location":"faq/#what-support-is-there-for-recent-pc-hardware-what-about-gpu","title":"What support is there for recent PC hardware ? What about GPU ?","text":"As OpenBLAS is a volunteer project, it can take some time for the combination of a capable developer, free time, and particular hardware to come along, even for relatively common processors. Starting from 0.3.1, support is being added for AVX 512 (TARGET=SKYLAKEX), requiring a compiler that is capable of handling avx512 intrinsics. While AMD Zen processors should be autodetected by the build system, as of 0.3.2 they are still handled exactly like Intel Haswell. There once was an effort to build an OpenCL implementation that one can still find at https://github.com/xianyi/clOpenBLAS , but work on this stopped in 2015.
"},{"location":"faq/#how-about-the-level-3-blas-performance-on-intel-sandy-bridge","title":"How about the level 3 BLAS performance on Intel Sandy Bridge?","text":"We obtained a performance comparable with Intel MKL that actually outperformed Intel MKL in some cases. Here is the result of the DGEMM subroutine's performance on Intel Core i5-2500K Windows 7 SP1 64-bit:
"},{"location":"faq/#os-and-compiler","title":"OS and Compiler","text":""},{"location":"faq/#how-can-i-call-an-openblas-function-in-microsoft-visual-studio","title":"How can I call an OpenBLAS function in Microsoft Visual Studio?","text":"Please read this page.
"},{"location":"faq/#how-can-i-use-cblas-and-lapacke-without-c99-complex-number-support-eg-in-visual-studio","title":"How can I use CBLAS and LAPACKE without C99 complex number support (e.g. in Visual Studio)?","text":"Zaheer has fixed this bug. You can now use the structure instead of C99 complex numbers. Please read this issue page for details.
This issue is for using LAPACKE in Visual Studio.
"},{"location":"faq/#i-get-a-segfault-with-multi-threading-on-linux-whats-wrong","title":"I get a SEGFAULT with multi-threading on Linux. What's wrong?","text":"This may be related to a bug in the Linux kernel 2.6.32 (?). Try applying the patch segaults.patch to disable mbind using
patch < segfaults.patch\nand see if the crashes persist. Note that this patch will lead to many compiler warnings.
"},{"location":"faq/#when-i-make-the-library-there-is-no-such-instruction-xgetbv-error-whats-wrong","title":"When I make the library, there is no such instruction: `xgetbv' error. What's wrong?","text":"Please use GCC 4.4 and later version. This version supports xgetbv instruction. If you use the library for Sandy Bridge with AVX instructions, you should use GCC 4.6 and later version.
On Mac OS X, please use Clang 3.1 and later version. For example, make CC=clang
For the compatibility with old compilers (GCC < 4.4), you can enable NO_AVX flag. For example, make NO_AVX=1
"},{"location":"faq/#my-build-fails-due-to-the-linker-error-multiple-definition-of-dlamc3_-what-is-the-problem","title":"My build fails due to the linker error \"multiple definition of `dlamc3_'\". What is the problem?","text":"This linker error occurs if GNU patch is missing or if our patch for LAPACK fails to apply.
Background: OpenBLAS implements optimized versions of some LAPACK functions, so we need to disable the reference versions. If this process fails we end with duplicated implementations of the same function.
"},{"location":"faq/#my-build-worked-fine-and-passed-all-tests-but-running-make-lapack-test-ends-with-segfaults","title":"My build worked fine and passed all tests, but runningmake lapack-test ends with segfaults","text":"Some of the LAPACK tests, notably in xeigtstz, try to allocate around 10MB on the stack. You may need to use ulimit -s to change the default limits on your system to allow this.
make lapack-test ends with a number of errors in the summary report","text":"The LAPACK tests were primarily created to test the validity of the Reference-LAPACK implementation, which is implemented in unoptimized, single-threaded Fortran code. This makes it very sensitive to small numerical deviations that can result from the use of specialized cpu instructions that combine multiplications and additions without intermediate rounding and storing to memory (FMA), or from changing the order of mathematical operations by splitting an original problem workload into smaller tasks that are solved in parallel. As a result, you may encounter a small number of errors in the \"numerical\" column of the summary table at the end of the make lapack-test run - this is usually nothing to worry about, and the exact number and distribution of errors among the four data types will often vary with the optimization flags you supplied to the compiler, or the cpu model for which you built OpenBLAS. Sporadic errors in the column labeled other are normally the sign of failed convergence of iterative diagonalizations for the same reasons just mentioned. A more detailed error report is stored in the file testing_results.txt - this should be consulted in case of doubt. Care should be taken if you encounter numerical errors in the hundreds, or other errors accompanied by the LAPACK error message \"on entry to function_name parameter X had an illegal value\" that signals a problem with argument passing between individual functions. (See also this issue in the issue tracker on github for additional discussion, examples and links)
You can define the OPENBLAS_MAIN_FREE or GOTOBLAS_MAIN_FREE environment variable to disable threading affinity on runtime. For example, before the running,
export OPENBLAS_MAIN_FREE=1\nAlternatively, you can disable affinity feature with enabling NO_AFFINITY=1 in Makefile.rule.
"},{"location":"faq/#how-to-solve-undefined-reference-errors-when-statically-linking-against-libopenblasa","title":"How to solve undefined reference errors when statically linking against libopenblas.a","text":"On Linux, if OpenBLAS was compiled with threading support (USE_THREAD=1 by default), custom programs statically linked against libopenblas.a should also link to the pthread library e.g.:
gcc -static -I/opt/OpenBLAS/include -L/opt/OpenBLAS/lib -o my_program my_program.c -lopenblas -lpthread\nFailing to add the -lpthread flag will cause errors such as:
/opt/OpenBLAS/libopenblas.a(memory.o): In function `_touch_memory':\nmemory.c:(.text+0x15): undefined reference to `pthread_mutex_lock'\nmemory.c:(.text+0x41): undefined reference to `pthread_mutex_unlock'\n/opt/OpenBLAS/libopenblas.a(memory.o): In function `openblas_fork_handler':\nmemory.c:(.text+0x440): undefined reference to `pthread_atfork'\n/opt/OpenBLAS/libopenblas.a(memory.o): In function `blas_memory_alloc':\nmemory.c:(.text+0x7a5): undefined reference to `pthread_mutex_lock'\nmemory.c:(.text+0x825): undefined reference to `pthread_mutex_unlock'\n/opt/OpenBLAS/libopenblas.a(memory.o): In function `blas_shutdown':\nmemory.c:(.text+0x9e1): undefined reference to `pthread_mutex_lock'\nmemory.c:(.text+0xa6e): undefined reference to `pthread_mutex_unlock'\n/opt/OpenBLAS/libopenblas.a(blas_server.o): In function `blas_thread_server':\nblas_server.c:(.text+0x273): undefined reference to `pthread_mutex_lock'\nblas_server.c:(.text+0x287): undefined reference to `pthread_mutex_unlock'\nblas_server.c:(.text+0x33f): undefined reference to `pthread_cond_wait'\n/opt/OpenBLAS/libopenblas.a(blas_server.o): In function `blas_thread_init':\nblas_server.c:(.text+0x416): undefined reference to `pthread_mutex_lock'\nblas_server.c:(.text+0x4be): undefined reference to `pthread_mutex_init'\nblas_server.c:(.text+0x4ca): undefined reference to `pthread_cond_init'\nblas_server.c:(.text+0x4e0): undefined reference to `pthread_create'\nblas_server.c:(.text+0x50f): undefined reference to `pthread_mutex_unlock'\n...\nThe -lpthread is not required when linking dynamically against libopenblas.so.0.
Minimum requirement to actually run AVX2-enabled software like OpenBLAS is kernel-2.6.32-358, shipped with EL6U4 in 2013
The binutils package from RHEL6 does not know the instruction vpermpd or any other AVX2 instruction. You can download a newer binutils package from Enterprise Linux software collections, following instructions here: https://www.softwarecollections.org/en/scls/rhscl/devtoolset-3/ After configuring repository you need to install devtoolset-?-binutils to get later usable binutils package
$ yum search devtoolset-\\?-binutils\n$ sudo yum install devtoolset-3-binutils\n$ scl --list\ndevtoolset-3\nrh-python35\n$ scl enable devtoolset-3 -- make DYNAMIC_ARCH=1\nBy default, QEMU reports the CPU as \"QEMU Virtual CPU version 2.2.0\", which shares CPUID with existing 32bit CPU even in 64bit virtual machine, and OpenBLAS recognizes it as PENTIUM2. Depending on the exact combination of CPU features the hypervisor choses to expose, this may not correspond to any CPU that exists, and OpenBLAS will error when trying to build. To fix this, pass -cpu host or -cpu passthough to QEMU, or another CPU model. Similarly, the XEN hypervisor may not pass through all features of the host cpu while reporting the cpu type itself correctly, which can lead to compiler error messages about an \"ABI change\" when compiling AVX512 code. Again changing the Xen configuration by running e.g. \"xen-cmdline --set-xen cpuid=avx512\" should get around this (as would building OpenBLAS for an older cpu lacking that particular feature, e.g. TARGET=HASWELL)
Trying to compile OpenBLAS with IBM XL ends with error messages about unknown register names\nlike \"vs32\". Working around these by using known alternate names for the vector registers only leads to another assembler error about unsupported constraints. This is a known deficiency in the IBM compiler at least up to and including 16.1.0 (and in the POWER version of clang, from which it is derived) - use gcc instead. (See issues #1078 and #1699 for related discussions)
"},{"location":"faq/#replacing-system-blasupdating-apt-openblas-in-mintubuntudebian","title":"Replacing system BLAS/updating APT OpenBLAS in Mint/Ubuntu/Debian","text":"Debian and Ubuntu LTS versions provide OpenBLAS package which is not updated after initial release, and under circumstances one might want to use more recent version of OpenBLAS e.g. to get support for newer CPUs
Ubuntu and Debian provides 'alternatives' mechanism to comfortably replace BLAS and LAPACK libraries systemwide.
After successful build of OpenBLAS (with DYNAMIC_ARCH set to 1)
$ make clean\n$ make DYNAMIC_ARCH=1\n$ sudo make DYNAMIC_ARCH=1 install\n$ sudo apt install libblas-dev liblapack-dev\n$ sudo update-alternatives --install /usr/lib/libblas.so.3 libblas.so.3 /opt/OpenBLAS/lib/libopenblas.so.0 41 \\\n --slave /usr/lib/liblapack.so.3 liblapack.so.3 /opt/OpenBLAS/lib/libopenblas.so.0\n$ sudo update-alternatives --remove libblas.so.3 /opt/OpenBLAS/lib/libopenblas.so.0\n$ update-alternatives --display libblas.so.3 to find out what layout your system has."},{"location":"faq/#i-built-openblas-for-use-with-some-other-software-but-that-software-cannot-find-it","title":"I built OpenBLAS for use with some other software, but that software cannot find it","text":"Openblas installs as a single library named libopenblas.so, while some programs may be searching for a separate libblas.so and liblapack.so so you may need to create appropriate symbolic links (ln -s libopenblas.so libblas.so; ln -s libopenblas.so liblapack.so) or copies. Also make sure that the installation location (usually /opt/OpenBLAS/lib or /usr/local/lib) is among the library search paths of your system.
You probably tried to include a cblas.h that you simply copied from the OpenBLAS source, instead you need to run make install after building OpenBLAS and then use the modified cblas.h that this step builds in the installation path (usually either /usr/local/include, /opt/OpenBLAS/include or whatever you specified as PREFIX= on the make install)
This is due to different interpretations of the (informal) standard for passing characters as arguments between C and FORTRAN functions. As the method for storing text differs in the two languages, when C calls Fortran the text length is passed as an \"invisible\" additional parameter. Historically, this has not been required when the text is just a single character, so older code like the Reference-LAPACK bundled with OpenBLAS does not do it. Recently gcc's checking has changed to require it, but there is no consensus yet if and how the existing LAPACK (and many other codebases) should adapt. (And for actual compilation, gcc has mostly backtracked and provided compatibility options - hence the default build settings in the OpenBLAS Makefiles add -fno-optimize-sibling-calls to the gfortran options to prevent miscompilation with \"affected\" versions. See ticket 2154 in the issue tracker for more details and links)
"},{"location":"faq/#build-fails-with-lots-of-errors-about-undefined-gemm_unroll_m","title":"Build fails with lots of errors about undefined ?GEMM_UNROLL_M","text":"Your cpu is apparently too new to be recognized by the build scripts, so they failed to assign appropriate parameters for the block algorithm. Do a make clean and try again with TARGET set to one of the cpu models listed in TargetList.txt - for x86_64 this will usually be HASWELL.
This is a limitation in the maximum length of a command on OSX, coupled with how CMAKE works. You should be able to work around this by adding the option -DCMAKE_Fortran_USE_RESPONSE_FILE_FOR_OBJECTS=1 to your CMAKE arguments.
There have been a few reports of wrong calculation results and build-time test failures when building in a container environment managed by the OSX version of Docker Desktop, which uses the xhyve virtualizer underneath. Judging from these reports, AVX2 support in xhyve appears to be subtly broken but a corresponding ticket in the xhyve issue tracker has not drawn any reaction or comment since 2019. Therefore it is strongly recommended to build OpenBLAS with the NO_AVX2=1 option when inside a container under (or for later use with) the Docker Desktop environment on Intel-based Apple hardware.
"},{"location":"faq/#usage","title":"Usage","text":""},{"location":"faq/#program-is-terminated-because-you-tried-to-allocate-too-many-memory-regions","title":"Program is Terminated. Because you tried to allocate too many memory regions","text":"In OpenBLAS, we mange a pool of memory buffers and allocate the number of buffers as the following.
#define NUM_BUFFERS (MAX_CPU_NUMBER * 2)\nPlease build OpenBLAS with larger NUM_THREADS. For example, make NUM_THREADS=32 or make NUM_THREADS=64. In Makefile.system, we will set MAX_CPU_NUMBER=NUM_THREADS.
The environment variable which control the kernel selection is OPENBLAS_CORETYPE (see driver/others/dynamic.c) e.g. export OPENBLAS_CORETYPE=Haswell. And the function char* openblas_get_corename() returns the used target.
This symbol gets defined only when OpenBLAS is built with \"make DYNAMIC_ARCH=1\" (which is what distributors will choose to ensure support for more than just one CPU type).
"},{"location":"faq/#how-can-i-find-out-at-runtime-what-options-the-library-was-built-with","title":"How can I find out at runtime what options the library was built with ?","text":"OpenBLAS has two utility functions that may come in here:
openblas_get_parallel() will return 0 for a single-threaded library, 1 if multithreading without OpenMP, 2 if built with USE_OPENMP=1
openblas_get_config() will return a string containing settings such as USE64BITINT or DYNAMIC_ARCH that were active at build time, as well as the target cpu (or in case of a dynamic_arch build, the currently detected one).
"},{"location":"faq/#after-making-openblas-i-find-that-the-static-library-is-multithreaded-but-the-dynamic-one-is-not","title":"After making OpenBLAS, I find that the static library is multithreaded, but the dynamic one is not ?","text":"The shared OpenBLAS library you built is probably working fine as well, but your program may be picking up a different (probably single-threaded) version from one of the standard system paths like /usr/lib on startup. Running ldd /path/to/your/program will tell you which library the linkage loader will actually use.
Specifying the \"correct\" library location with the -L flag (like -L /opt/OpenBLAS/lib) when linking your program only defines which library will be used to see if all symbols can be resolved, you will need to add an rpath entry to the binary (using -Wl,rpath=/opt/OpenBLAS/lib) to make it request searching that location. Alternatively, remove the \"wrong old\" library (if you can), or set LD_LIBRARY_PATH to the desired location before running your program.
You need to edit file src/cuda/cuda_dgemm.c in the NVIDIA version of HPL, change the \"handle2\" and \"handle\" dlopen calls to use libopenblas.so instead of libmkl_intel_lp64.so, and add an trailing underscore in the dlsym lines for dgemm_mkl and dtrsm_mkl (like dgemm_mkl = (void(*)())dlsym(handle, \u201cdgemm_\u201d);)
The power saving mechanisms of your board may have shut down some cores, making them invisible to OpenBLAS in its startup phase. Try bringing them online before starting your calculation.
"},{"location":"faq/#speed-varies-wildly-between-individual-runs-on-a-typical-armv8-smartphone-processor","title":"Speed varies wildly between individual runs on a typical ARMV8 smartphone processor","text":"Check the technical specifications, it could be that the SoC combines fast and slow cpus and threads can end up on either. In that case, binding the process to specific cores e.g. by setting OMP_PLACES=cores may help. (You may need to experiment with OpenMP options, it has been reported that using OMP_NUM_THREADS=2 OMP_PLACES=cores caused a huge drop in performance on a 4+4 core chip while OMP_NUM_THREADS=2 OMP_PLACES=cores(2) worked as intended - as did OMP_PLACES=cores with 4 threads)
Multithreading support in OpenBLAS requires the use of internal buffers for sharing partial results, the number and size of which is defined at compile time. Unless you specify NUM_THREADS in your make or cmake command, the build scripts try to autodetect the number of cores available in your build host to size the library to match. This unfortunately means that if you move the resulting binary from a small \"front-end node\" to a larger \"compute node\" later, it will still be limited to the hardware capabilities of the original system. The solution is to set NUM_THREADS to a number big enough to encompass the biggest systems you expect to run the binary on - at runtime, it will scale down the maximum number of threads it uses to match the number of cores physically available.
"},{"location":"faq/#getting-elf-load-command-addressoffset-not-properly-aligned-when-loading-libopenblasso","title":"Getting \"ELF load command address/offset not properly aligned\" when loading libopenblas.so","text":"If you get a message \"error while loading shared libraries: libopenblas.so.0: ELF load command address/offset not properly aligned\" when starting a program that is (dynamically) linked to OpenBLAS, this is very likely due to a bug in the GNU linker (ld) that is part of the GNU binutils package. This error was specifically observed on older versions of Ubuntu Linux updated with the (at the time) most recent binutils version 2.38, but an internet search turned up sporadic reports involving various other libraries dating back several years. A bugfix was created by the binutils developers and should be available in later versions of binutils.(See issue 3708 for details)
"},{"location":"faq/#using-openblas-with-openmp","title":"Using OpenBLAS with OpenMP","text":"OpenMP provides its own locking mechanisms, so when your code makes BLAS/LAPACK calls from inside OpenMP parallel regions it is imperative that you use an OpenBLAS that is built with USE_OPENMP=1, as otherwise deadlocks might occur. Furthermore, OpenBLAS will automatically restrict itself to using only a single thread when called from an OpenMP parallel region. When it is certain that calls will only occur from the main thread of your program (i.e. outside of omp parallel constructs), a standard pthreads build of OpenBLAS can be used as well. In that case it may be useful to tune the linger behaviour of idle threads in both your OpenMP program (e.g. set OMP_WAIT_POLICY=passive) and OpenBLAS (by redefining the THREAD_TIMEOUT variable at build time, or setting the environment variable OPENBLAS_THREAD_TIMEOUT smaller than the default 26) so that the two alternating thread pools do not unnecessarily hog the cpu during the handover.
"},{"location":"install/","title":"Install OpenBLAS","text":"OpenBLAS can be installed through package managers or from source. If you only want to use OpenBLAS rather than make changes to it, we recommend installing a pre-built binary package with your package manager of choice.
This page contains an overview of installing with package managers as well as from source. For the latter, see further down on this page.
"},{"location":"install/#installing-with-a-package-manager","title":"Installing with a package manager","text":"Note
Almost every package manager provides OpenBLAS packages; the list on this page is not comprehensive. If your package manager of choice isn't shown here, please search its package database for openblas or libopenblas.
On Linux, OpenBLAS can be installed with the system package manager, or with a package manager like Conda (or alternative package managers for the conda-forge ecosystem, like Mamba, Micromamba, or Pixi), Spack, or Nix. For the latter set of tools, the package name in all cases is openblas. Since package management in quite a few of these tools is declarative (i.e., managed by adding openblas to a metadata file describing the dependencies for your project or environment), we won't attempt to give detailed instructions for these tools here.
Linux distributions typically split OpenBLAS up in two packages: one containing the library itself (typically named openblas or libopenblas), and one containing headers, pkg-config and CMake files (typically named the same as the package for the library with -dev or -devel appended; e.g., openblas-devel). Please keep in mind that if you want to install OpenBLAS in order to use it directly in your own project, you will need to install both of those packages.
Distro-specific installation commands:
Debian/Ubuntu/Mint/KaliopenSUSE/SLEFedora/CentOS/RHELArch/Manjaro/Antergos$ sudo apt update\n$ sudo apt install libopenblas-dev\nupdate-alternatives mechanism: $ sudo update-alternatives --config libblas.so.3\n$ sudo zypper refresh\n$ sudo zypper install openblas-devel\nOpenBLAS can be configured as the default BLAS through the update-alternatives mechanism:
$ sudo update-alternatives --config libblas.so.3\n$ dnf check-update\n$ dnf install openblas-devel\nWarning
Fedora does not ship the pkg-config files for OpenBLAS. Instead, it wants you to link against FlexiBLAS (which uses OpenBLAS by default as its backend on Fedora), which you can install with:
$ dnf install flexiblas-devel\nFor CentOS and RHEL, OpenBLAS packages are provided via the Fedora EPEL repository. After adding that repository and its repository keys, you can install openblas-devel with either dnf or yum.
$ sudo pacman -S openblas\nOpenBLAS can be installed with conda (or mamba, micromamba, or pixi) from conda-forge:
conda install openblas\nConda-forge provides a method for switching the default BLAS implementation used by all packages. To use that for OpenBLAS, install libblas=*=*openblas (see the docs on this mechanism for more details).
OpenBLAS can be installed with vcpkg:
# In classic mode:\nvcpkg install openblas\n\n# Or in manifest mode:\nvcpkg add port openblas\nWindows is the only platform for which binaries are made available by the OpenBLAS project itself. They can be downloaded from the GitHub Releases](https://github.com/OpenMathLib/OpenBLAS/releases) page. These binaries are built with MinGW, using the following build options:
NUM_THREADS=64 TARGET=GENERIC DYNAMIC_ARCH=1 DYNAMIC_OLDER=1 CONSISTENT_FPCSR=1 INTERFACE=0\n.zip bundle there (note that you will need to edit the provided openblas.pc and OpenBLASConfig.cmake to reflect the installation path on your computer, as distributed they have \"win\" or \"win64\" reflecting the local paths on the system they were built on). Note that the same binaries can be downloaded from SourceForge; this is mostly of historical interest.
"},{"location":"install/#macos","title":"macOS","text":"To install OpenBLAS with a package manager on macOS, run:
HomebrewMacPortsConda-forge% brew install openblas\n% sudo port install OpenBLAS-devel\n% conda install openblas\nConda-forge provides a method for switching the default BLAS implementation used by all packages. To use that for OpenBLAS, install libblas=*=*openblas (see the docs on this mechanism for more details).
You can install OpenBLAS from the FreeBSD Ports collection:
pkg install openblas\nWe recommend download the latest stable version from the GitHub Releases page, or checking it out from a git tag, rather than a dev version from the develop branch.
Tip
The User manual contains a section with detailed information on compiling OpenBLAS, including how to customize builds and how to cross-compile. Please read that documentation first. This page contains only platform-specific build information, and assumes you already understand the general build system invocations to build OpenBLAS, with the specific build options you want to control multi-threading and other non-platform-specific behavior).
"},{"location":"install/#linux-and-macos","title":"Linux and macOS","text":"Ensure you have C and Fortran compilers installed, then simply type make to compile the library. There are no other build dependencies, nor unusual platform-specific environment variables to set or other system setup to do.
Note
When building in an emulator (KVM, QEMU, etc.), please make sure that the combination of CPU features exposed to the virtual environment matches that of an existing CPU to allow detection of the CPU model to succeed. (With qemu, this can be done by passing -cpu host or a supported model name at invocation).
We support building OpenBLAS with either MinGW or Visual Studio on Windows. Using MSVC will yield an OpenBLAS build with the Windows platform-native ABI. Using MinGW will yield a different ABI. We'll describe both methods in detail in this section, since the process for each is quite different.
"},{"location":"install/#visual-studio-native-windows-abi","title":"Visual Studio & native Windows ABI","text":"For Visual Studio, you can use CMake to generate Visual Studio solution files; note that you will need at least CMake 3.11 for linking to work correctly).
Note that you need a Fortran compiler if you plan to build and use the LAPACK functions included with OpenBLAS. The sections below describe using either flang as an add-on to clang/LLVM or gfortran as part of MinGW for this purpose. If you want to use the Intel Fortran compiler (ifort or ifx) for this, be sure to also use the Intel C compiler (icc or icx) for building the C parts, as the ABI imposed by ifort is incompatible with MSVC
A fully-optimized OpenBLAS that can be statically or dynamically linked to your application can currently be built for the 64-bit architecture with the LLVM compiler infrastructure. We're going to use Miniconda3 to grab all of the tools we need, since some of them are in an experimental status. Before you begin, you'll need to have Microsoft Visual Studio 2015 or newer installed.
winget install --id Anaconda.Miniconda3, or easily download from conda.io.%USERPROFILE%\\miniconda3\\shell\\condabin\\conda-hook.ps1.cd to change to the directory where you want to build OpenBLAS.conda update -n base conda\nconda config --add channels conda-forge\nconda install -y cmake flang clangdev perl libflang ninja\nStill in the Anaconda Command Prompt window, activate the 64-bit MSVC environment with vcvarsall x64. On Windows 11 with Visual Studio 2022, this would be done by invoking:
\"c:\\Program Files\\Microsoft Visual Studio\\2022\\Community\\vc\\Auxiliary\\Build\\vcvars64.bat\"\nWith VS2019, the command should be the same (except for the year number of course). For other versions of MSVC, please check the Visual Studio documentation for exactly how to invoke the vcvars64.bat script.
Confirm that the environment is active by typing link. This should return a long list of possible options for the link command. If it just returns \"command not found\" or similar, review and retype the call to vcvars64.bat.
Note
if you are working from a Visual Studio command prompt window instead (so that you do not have to do the vcvars call), you need to invoke conda activate so that CONDA_PREFIX etc. get set up correctly before proceeding to step 6. Failing to do so will lead to link errors like libflangmain.lib not getting found later in the build.
Now configure the project with CMake. Starting in the project directory, execute the following:
set \"LIB=%CONDA_PREFIX%\\Library\\lib;%LIB%\"\nset \"CPATH=%CONDA_PREFIX%\\Library\\include;%CPATH%\"\nmkdir build\ncd build\ncmake .. -G \"Ninja\" -DCMAKE_CXX_COMPILER=clang-cl -DCMAKE_C_COMPILER=clang-cl -DCMAKE_Fortran_COMPILER=flang -DCMAKE_MT=mt -DBUILD_WITHOUT_LAPACK=no -DNOFORTRAN=0 -DDYNAMIC_ARCH=ON -DCMAKE_BUILD_TYPE=Release\nYou may want to add further options in the cmake command here. For instance, the default only produces a static .lib version of the library. If you would rather have a DLL, add -DBUILD_SHARED_LIBS=ON above. Note that this step only creates some command files and directories, the actual build happens next.
Build the project:
cmake --build . --config Release\nlib directory, and various build-time tests in the test, ctest and openblas_utest directories. However it will not separate the header files you might need for building your own programs from those used internally. To put all relevant files in a more convenient arrangement, run the next step. Install all relevant files created by the build:
cmake --install . --prefix c:\\opt -v\nC:\\opt as given above, this would be: C:\\opt\\include\\openblas for the header files, C:\\opt\\bin for the libopenblas.dll shared library,C:\\opt\\lib for the static library, andC:\\opt\\share holds various support files that enable other cmake-based build scripts to find OpenBLAS automatically.Change in complex types for Visual Studio 2017 and up
In newer Visual Studio versions, Microsoft has changed how it handles complex types. Even when using a precompiled version of OpenBLAS, you might need to define LAPACK_COMPLEX_CUSTOM in order to define complex types properly for MSVC. For example, some variant of the following might help:
#if defined(_MSC_VER)\n #include <complex.h>\n #define LAPACK_COMPLEX_CUSTOM\n #define lapack_complex_float _Fcomplex\n #define lapack_complex_double _Dcomplex\n#endif\nFor reference, see openblas#3661, lapack#683, and this Stack Overflow question.
Building 32-bit binaries with MSVC
This method may produce binaries which demonstrate significantly lower performance than those built with the other methods. The Visual Studio compiler does not support the dialect of assembly used in the cpu-specific optimized files, so only the \"generic\" TARGET which is written in pure C will get built. For the same reason it is not possible (and not necessary) to use -DDYNAMIC_ARCH=ON in a Visual Studio build. You may consider building for the 32-bit architecture using the GNU (MinGW) ABI instead.
To generate Visual Studio solution files, ensure CMake is installed and then run:
# Do this from Powershell so cmake can find visual studio\ncmake -G \"Visual Studio 14 Win64\" -DCMAKE_BUILD_TYPE=Release .\nTo then build OpenBLAS using those solution files from within Visual Studio, we also need Perl. Please install it and ensure it's on the PATH (see, e.g., this Stack Overflow question for how).
If you build from within Visual Studio, the dependencies may not be automatically configured: if you try to build libopenblas directly, it may fail with a message saying that some .obj files aren't found. If this happens, you can work around the problem by building the projects that libopenblas depends on before building libopenblas itself.
OpenBLAS can be built targeting Universal Windows Platform (UWP) like this:
Remove the generated CMakeCache.txt and the CMakeFiles directory from the OpenBLAS source directory, then re-run CMake with the following options:
# do this to build UWP compatible solution files\ncmake -G \"Visual Studio 14 Win64\" -DCMAKE_SYSTEM_NAME=WindowsStore -DCMAKE_SYSTEM_VERSION=\"10.0\" -DCMAKE_SYSTEM_PROCESSOR=AMD64 -DVS_WINRT_COMPONENT=TRUE -DCMAKE_BUILD_TYPE=Release .\nNote
The resulting library from building with MinGW as described below can be used in Visual Studio, but it can only be linked dynamically. This configuration has not been thoroughly tested and should be considered experimental.
To build OpenBLAS on Windows with MinGW:
gfortran package as well (unless you really want to build the BLAS part of OpenBLAS only) and check that gcc and gfortran are the same version. In addition, please install MSYS2 with MinGW.make. OpenBLAS will detect the compiler and CPU automatically. libopenblas.a and the shared library libopenblas.dll in the folder. You can type make PREFIX=/your/installation/path install to install the library to a certain location.Note that OpenBLAS will generate the import library libopenblas.dll.a for libopenblas.dll by default.
If you want to generate Windows-native PDB files from a MinGW build, you can use the cv2pdb tool to do so.
To then use the built OpenBLAS shared library in Visual Studio:
OPENBLAS_TOP_DIR/libopenblas.dll.a) and the shared library (libopenblas.dll) into the same folder (this must be the folder of your project that is going to use the BLAS library. You may need to add libopenblas.dll.a to the linker input list: properties->Linker->Input).cblas.h in /your/installation/path/include in Visual Studio. Please see openblas#95 for more details.Limitations of using the MinGW build within Visual Studio
A fully functional native OpenBLAS for WoA that can be built as both a static and dynamic library using LLVM toolchain and Visual Studio 2022. Before starting to build, make sure that you have installed Visual Studio 2022 on your ARM device, including the \"Desktop Development with C++\" component (that contains the cmake tool). (Note that you can use the free \"Visual Studio 2022 Community Edition\" for this task. In principle it would be possible to build with VisualStudio alone, but using the LLVM toolchain enables native compilation of the Fortran sources of LAPACK and of all the optimized assembly files, which VisualStudio cannot handle on its own)
Clone OpenBLAS to your local machine and checkout to latest release of OpenBLAS (unless you want to build the latest development snapshot - here we are using the 0.3.28 release as the example, of course this exact version may be outdated by the time you read this)
git clone https://github.com/OpenMathLib/OpenBLAS.git\ncd OpenBLAS\ngit checkout v0.3.28\nInstall Latest LLVM toolchain for WoA:
Download the Latest LLVM toolchain for WoA from the Release page. At the time of writing, this is version 19.1.5 - be sure to select the latest release for which you can find a precompiled package whose name ends in \"-woa64.exe\" (precompiled packages usually lag a week or two behind their corresponding source release). Make sure to enable the option \u201cAdd LLVM to the system PATH for all the users\u201d Note: Make sure that the path of LLVM toolchain is at the top of Environment Variables section to avoid conflicts between the set of compilers available in the system path
From the start menu search for \u201cARM64 Native Tools Command Prompt for Visual Studio 2022\u201d Alternatively open command prompt, run the following command to activate the environment: \"C:\\Program Files\\Microsoft Visual Studio\\2022\\Community\\VC\\Auxiliary\\Build\\vcvarsarm64.bat\"
Navigate to the OpenBLAS source code directory and start building OpenBLAS by invoking Ninja:
```cmd\n cd OpenBLAS\n mkdir build\n cd build\n\n cmake .. -G Ninja -DCMAKE_BUILD_TYPE=Release -DTARGET=ARMV8 -DBINARY=64 -DCMAKE_C_COMPILER=clang-cl -DCMAKE_C_COMPILER=arm64-pc-windows-msvc -DCMAKE_ASM_COMPILER=arm64-pc-windows-msvc -DCMAKE_Fortran_COMPILER=flang-new\n\n ninja -j16\n ```\nNote: You might want to include additional options in the cmake command here. For example, the default configuration only generates a static.lib version of the library. If you prefer a DLL, you can add -DBUILD_SHARED_LIBS=ON.
Note that it is also possible to use the same setup to build OpenBLAS with Make, if you prepare Makefiles over the CMake build for some reason:
```cmd\n$ make CC=clang-cl FC=flang-new AR=\"llvm-ar\" TARGET=ARMV8 ARCH=arm64 RANLIB=\"llvm-ranlib\" MAKE=make\n```\nMicrosoft Windows has this thing called \"import libraries\". You need it for MSVC; you don't need it for MinGW because the ld linker is smart enough - however, you may still want it for some reason, so we'll describe the process for both MSVC and MinGW.
Import libraries are compiled from a list of what symbols to use, which are contained in a .def file. A .def file should be already be present in the exports directory under the top-level OpenBLAS directory after you've run a build. In your shell, move to this directory: cd exports.
Unlike MinGW, MSVC absolutely requires an import library. Now the C ABI of MSVC and MinGW are actually identical, so linking is actually okay (any incompatibility in the C ABI would be a bug).
The import libraries of MSVC have the suffix .lib. They are generated from a .def file using MSVC's lib.exe. See the MSVC instructions.
MinGW import libraries have the suffix .a, just like static libraries. Our goal is to produce the file libopenblas.dll.a.
You need to first insert a line LIBRARY libopenblas.dll in libopenblas.def:
cat <(echo \"LIBRARY libopenblas.dll\") libopenblas.def > libopenblas.def.1\nmv libopenblas.def.1 libopenblas.def\nNow the .def file probably looks like:
LIBRARY libopenblas.dll\nEXPORTS\n caxpy=caxpy_ @1\n caxpy_=caxpy_ @2\n ...\ndlltool -d libopenblas.def -l libopenblas.dll.a Again, there is basically no point in making an import library for use in MinGW. It actually slows down linking.
"},{"location":"install/#android","title":"Android","text":"To build OpenBLAS for Android, you will need the following tools installed on your machine:
The next two sections below describe how to build with Clang for ARMV7 and ARMV8 targets, respectively. The same basic principles as described below for ARMV8 should also apply to building an x86 or x86-64 version (substitute something like NEHALEM for the target instead of ARMV8, and replace all the aarch64 in the toolchain paths with x86 or x96_64 as appropriate).
Historic note
Since NDK version 19, the default toolchain is provided as a standalone toolchain, so building one yourself following building a standalone toolchain should no longer be necessary.
"},{"location":"install/#building-for-armv7","title":"Building for ARMV7","text":"# Set path to ndk-bundle\nexport NDK_BUNDLE_DIR=/path/to/ndk-bundle\n\n# Set the PATH to contain paths to clang and arm-linux-androideabi-* utilities\nexport PATH=${NDK_BUNDLE_DIR}/toolchains/arm-linux-androideabi-4.9/prebuilt/linux-x86_64/bin:${NDK_BUNDLE_DIR}/toolchains/llvm/prebuilt/linux-x86_64/bin:$PATH\n\n# Set LDFLAGS so that the linker finds the appropriate libgcc\nexport LDFLAGS=\"-L${NDK_BUNDLE_DIR}/toolchains/arm-linux-androideabi-4.9/prebuilt/linux-x86_64/lib/gcc/arm-linux-androideabi/4.9.x\"\n\n# Set the clang cross compile flags\nexport CLANG_FLAGS=\"-target arm-linux-androideabi -marm -mfpu=vfp -mfloat-abi=softfp --sysroot ${NDK_BUNDLE_DIR}/platforms/android-23/arch-arm -gcc-toolchain ${NDK_BUNDLE_DIR}/toolchains/arm-linux-androideabi-4.9/prebuilt/linux-x86_64/\"\n\n#OpenBLAS Compile\nmake TARGET=ARMV7 ONLY_CBLAS=1 AR=ar CC=\"clang ${CLANG_FLAGS}\" HOSTCC=gcc ARM_SOFTFP_ABI=1 -j4\nOn macOS, it may also be necessary to give the complete path to the ar utility in the make command above, like so:
AR=${NDK_BUNDLE_DIR}/toolchains/arm-linux-androideabi-4.9/prebuilt/darwin-x86_64/bin/arm-linux-androideabi-gcc-ar\nmalformed archive header name at 8 when the native macOS ar command was invoked instead. Note that with recent NDK versions, the AR tool may be named llvm-ar rather than what is assumed above."},{"location":"install/#building-for-armv8","title":"Building for ARMV8","text":"# Set path to ndk-bundle\nexport NDK_BUNDLE_DIR=/path/to/ndk-bundle/\n\n# Export PATH to contain directories of clang and aarch64-linux-android-* utilities\nexport PATH=${NDK_BUNDLE_DIR}/toolchains/aarch64-linux-android-4.9/prebuilt/linux-x86_64/bin/:${NDK_BUNDLE_DIR}/toolchains/llvm/prebuilt/linux-x86_64/bin:$PATH\n\n# Setup LDFLAGS so that loader can find libgcc and pass -lm for sqrt\nexport LDFLAGS=\"-L${NDK_BUNDLE_DIR}/toolchains/aarch64-linux-android-4.9/prebuilt/linux-x86_64/lib/gcc/aarch64-linux-android/4.9.x -lm\"\n\n# Setup the clang cross compile options\nexport CLANG_FLAGS=\"-target aarch64-linux-android --sysroot ${NDK_BUNDLE_DIR}/platforms/android-23/arch-arm64 -gcc-toolchain ${NDK_BUNDLE_DIR}/toolchains/aarch64-linux-android-4.9/prebuilt/linux-x86_64/\"\n\n# Compile\nmake TARGET=ARMV8 ONLY_CBLAS=1 AR=ar CC=\"clang ${CLANG_FLAGS}\" HOSTCC=gcc -j4\nTARGET=CORTEXA57 in place of ARMV8 will pick up better optimized routines. Implementations for the CORTEXA57 target are compatible with all other ARMV8 targets. Note: for NDK 23b, something as simple as:
export PATH=/opt/android-ndk-r23b/toolchains/llvm/prebuilt/linux-x86_64/bin/:$PATH\nmake HOSTCC=gcc CC=/opt/android-ndk-r23b/toolchains/llvm/prebuilt/linux-x86_64/bin/aarch64-linux-android31-clang ONLY_CBLAS=1 TARGET=ARMV8\nllvm-ar) is also necessary. Alternative build script for 3 architectures This script will build OpenBLAS for 3 architecture (ARMV7, ARMV8, X86) and install them to /opt/OpenBLAS/lib. Of course you can also copy only the section that is of interest to you - also notice that the AR= line may need adapting to the name of the ar tool provided in your $TOOLCHAIN/bin - for example llvm-ar in some recent NDK versions. It was tested on macOS with NDK version 21.3.6528147.
export NDK=YOUR_PATH_TO_SDK/Android/sdk/ndk/21.3.6528147\nexport TOOLCHAIN=$NDK/toolchains/llvm/prebuilt/darwin-x86_64\n\nmake clean\nmake \\\n TARGET=ARMV7 \\\n ONLY_CBLAS=1 \\\n CC=\"$TOOLCHAIN\"/bin/armv7a-linux-androideabi21-clang \\\n AR=\"$TOOLCHAIN\"/bin/arm-linux-androideabi-ar \\\n HOSTCC=gcc \\\n ARM_SOFTFP_ABI=1 \\\n -j4\nsudo make install\n\nmake clean\nmake \\\n TARGET=CORTEXA57 \\\n ONLY_CBLAS=1 \\\n CC=$TOOLCHAIN/bin/aarch64-linux-android21-clang \\\n AR=$TOOLCHAIN/bin/aarch64-linux-android-ar \\\n HOSTCC=gcc \\\n-j4\nsudo make install\n\nmake clean\nmake \\\n TARGET=ATOM \\\n ONLY_CBLAS=1 \\\nCC=\"$TOOLCHAIN\"/bin/i686-linux-android21-clang \\\nAR=\"$TOOLCHAIN\"/bin/i686-linux-android-ar \\\nHOSTCC=gcc \\\nARM_SOFTFP_ABI=1 \\\n-j4\nsudo make install\n\n## This will build for x86_64 \nmake clean\nmake \\\n TARGET=ATOM BINARY=64\\\nONLY_CBLAS=1 \\\nCC=\"$TOOLCHAIN\"/bin/x86_64-linux-android21-clang \\\nAR=\"$TOOLCHAIN\"/bin/x86_64-linux-android-ar \\\nHOSTCC=gcc \\\nARM_SOFTFP_ABI=1 \\\n-j4\nsudo make install\nAs none of the current developers uses iOS, the following instructions are what was found to work in our Azure CI setup, but as far as we know this builds a fully working OpenBLAS for this platform.
Go to the directory where you unpacked OpenBLAS,and enter the following commands:
CC=/Applications/Xcode_12.4.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin/clang\n\nCFLAGS= -O2 -Wno-macro-redefined -isysroot /Applications/Xcode_12.4.app/Contents/Developer/Platforms/iPhoneOS.platform/Developer/SDKs/iPhoneOS14.4.sdk -arch arm64 -miphoneos-version-min=10.0\n\nmake TARGET=ARMV8 DYNAMIC_ARCH=1 NUM_THREADS=32 HOSTCC=clang NOFORTRAN=1\nMIN_IOS_VERSION as necessary for your installation. E.g., change the version number to the minimum iOS version you want to target and execute this file to build the library."},{"location":"install/#mips","title":"MIPS","text":"For MIPS targets you will need latest toolchains:
You can use following commandlines for builds:
IMG_TOOLCHAIN_DIR={full IMG GNU/Linux Toolchain path including \"bin\" directory -- for example, /opt/linux_toolchain/bin}\nIMG_GCC_PREFIX=mips-img-linux-gnu\nIMG_TOOLCHAIN=${IMG_TOOLCHAIN_DIR}/${IMG_GCC_PREFIX}\n\n# I6400 Build (n32):\nmake BINARY=32 BINARY32=1 CC=$IMG_TOOLCHAIN-gcc AR=$IMG_TOOLCHAIN-ar FC=\"$IMG_TOOLCHAIN-gfortran -EL -mabi=n32\" RANLIB=$IMG_TOOLCHAIN-ranlib HOSTCC=gcc CFLAGS=\"-EL\" FFLAGS=$CFLAGS LDFLAGS=$CFLAGS TARGET=I6400\n\n# I6400 Build (n64):\nmake BINARY=64 BINARY64=1 CC=$IMG_TOOLCHAIN-gcc AR=$IMG_TOOLCHAIN-ar FC=\"$IMG_TOOLCHAIN-gfortran -EL\" RANLIB=$IMG_TOOLCHAIN-ranlib HOSTCC=gcc CFLAGS=\"-EL\" FFLAGS=$CFLAGS LDFLAGS=$CFLAGS TARGET=I6400\n\n# P6600 Build (n32):\nmake BINARY=32 BINARY32=1 CC=$IMG_TOOLCHAIN-gcc AR=$IMG_TOOLCHAIN-ar FC=\"$IMG_TOOLCHAIN-gfortran -EL -mabi=n32\" RANLIB=$IMG_TOOLCHAIN-ranlib HOSTCC=gcc CFLAGS=\"-EL\" FFLAGS=$CFLAGS LDFLAGS=$CFLAGS TARGET=P6600\n\n# P6600 Build (n64):\nmake BINARY=64 BINARY64=1 CC=$IMG_TOOLCHAIN-gcc AR=$IMG_TOOLCHAIN-ar FC=\"$IMG_TOOLCHAIN-gfortran -EL\" RANLIB=$IMG_TOOLCHAIN-ranlib HOSTCC=gcc CFLAGS=\"-EL\" FFLAGS=\"$CFLAGS\" LDFLAGS=\"$CFLAGS\" TARGET=P6600\n\nMTI_TOOLCHAIN_DIR={full MTI GNU/Linux Toolchain path including \"bin\" directory -- for example, /opt/linux_toolchain/bin}\nMTI_GCC_PREFIX=mips-mti-linux-gnu\nMTI_TOOLCHAIN=${IMG_TOOLCHAIN_DIR}/${IMG_GCC_PREFIX}\n\n# P5600 Build:\n\nmake BINARY=32 BINARY32=1 CC=$MTI_TOOLCHAIN-gcc AR=$MTI_TOOLCHAIN-ar FC=\"$MTI_TOOLCHAIN-gfortran -EL\" RANLIB=$MTI_TOOLCHAIN-ranlib HOSTCC=gcc CFLAGS=\"-EL\" FFLAGS=$CFLAGS LDFLAGS=$CFLAGS TARGET=P5600\nYou will need to install the following tools from the FreeBSD ports tree:
To compile run the command:
$ gmake CC=gcc FC=gfortran\nCortex-M is a widely used microcontroller that is present in a variety of industrial and consumer electronics. A common variant of the Cortex-M is the STM32F4xx series. Here, we will give instructions for building for that series.
First, install the embedded Arm GCC compiler from the Arm website. Then, create the following toolchain.cmake file:
set(CMAKE_SYSTEM_NAME Generic)\nset(CMAKE_SYSTEM_PROCESSOR arm)\n\nset(CMAKE_C_COMPILER \"arm-none-eabi-gcc.exe\")\nset(CMAKE_CXX_COMPILER \"arm-none-eabi-g++.exe\")\n\nset(CMAKE_EXE_LINKER_FLAGS \"--specs=nosys.specs\" CACHE INTERNAL \"\")\n\nset(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM NEVER)\nset(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY ONLY)\nset(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)\nset(CMAKE_FIND_ROOT_PATH_MODE_PACKAGE ONLY)\nThen build OpenBLAS with:
$ cmake .. -G Ninja -DCMAKE_C_COMPILER=arm-none-eabi-gcc -DCMAKE_TOOLCHAIN_FILE:PATH=\"toolchain.cmake\" -DNOFORTRAN=1 -DTARGET=ARMV5 -DEMBEDDED=1\nIn your embedded application, the following functions need to be provided for OpenBLAS to work correctly:
void free(void* ptr);\nvoid* malloc(size_t size);\nNote
If you are developing for an embedded platform, it is your responsibility to make sure that the device has sufficient memory for malloc calls. Libmemory provides one implementation of malloc for embedded platforms.
If the OpenBLAS DLLs are not linked correctly, you may see an error like \"The application was unable to start correctly (0xc000007b)\", which typically indicates a mismatch between 32-bit and 64-bit libraries.\u00a0\u21a9
This user manual covers compiling OpenBLAS itself, linking your code to OpenBLAS, example code to use the C (CBLAS) and Fortran (BLAS) APIs, and some troubleshooting tips. Compiling OpenBLAS is optional, since you may be able to install with a package manager.
Note
The OpenBLAS documentation does not contain API reference documentation for BLAS or LAPACK, since these are standardized APIs, the documentation for which can be found in other places. If you want to understand every BLAS and LAPACK function and definition, we recommend reading the Netlib BLAS and Netlib LAPACK documentation.
OpenBLAS does contain a limited number of functions that are non-standard, these are documented at OpenBLAS extension functions.
"},{"location":"user_manual/#compiling-openblas","title":"Compiling OpenBLAS","text":""},{"location":"user_manual/#normal-compile","title":"Normal compile","text":"The default way to build and install OpenBLAS from source is with Make:
make # add `-j4` to compile in parallel with 4 processes\nmake install\nBy default, the CPU architecture is detected automatically when invoking make, and the build is optimized for the detected CPU. To override the autodetection, use the TARGET flag:
# `make TARGET=xxx` sets target CPU: e.g. for an Intel Nehalem CPU:\nmake TARGET=NEHALEM\nTargetList.txt in the root of the repository."},{"location":"user_manual/#cross-compile","title":"Cross compile","text":"For a basic cross-compilation with Make, three steps need to be taken:
CC and FC environment variables to select the cross toolchains for C and Fortran.HOSTCC environment variable to select the host C compiler (i.e. the regular C compiler for the machine on which you are invoking the build).TARGET explicitly to the CPU architecture on which the produced OpenBLAS binaries will be used.Compile the library for ARM Cortex-A9 linux on an x86-64 machine (note: install only gnueabihf versions of the cross toolchain - see this issue comment for why):
make CC=arm-linux-gnueabihf-gcc FC=arm-linux-gnueabihf-gfortran HOSTCC=gcc TARGET=CORTEXA9\nCompile OpenBLAS for a loongson3a CPU on an x86-64 machine:
make BINARY=64 CC=mips64el-unknown-linux-gnu-gcc FC=mips64el-unknown-linux-gnu-gfortran HOSTCC=gcc TARGET=LOONGSON3A\nCompile OpenBLAS for loongson3a CPU with the loongcc (based on Open64) compiler on an x86-64 machine:
make CC=loongcc FC=loongf95 HOSTCC=gcc TARGET=LOONGSON3A CROSS=1 CROSS_SUFFIX=mips64el-st-linux-gnu- NO_LAPACKE=1 NO_SHARED=1 BINARY=32\nAdd DEBUG=1 to your build command, e.g.:
make DEBUG=1\nNote
Installing to a directory is optional; it is also possible to use the shared or static libraries directly from the build directory.
Use make install with the PREFIX flag to install to a specific directory:
make install PREFIX=/path/to/installation/directory\nThe default directory is /opt/OpenBLAS.
Important
Note that any flags passed to make during build should also be passed to make install to circumvent any install errors, i.e. some headers not being copied over correctly.
For more detailed information on building/installing from source, please read the Installation Guide.
"},{"location":"user_manual/#linking-to-openblas","title":"Linking to OpenBLAS","text":"OpenBLAS can be used as a shared or a static library.
"},{"location":"user_manual/#link-a-shared-library","title":"Link a shared library","text":"The shared library is normally called libopenblas.so, but not that the name may be different as a result of build flags used or naming choices by a distro packager (see [distributing.md] for details). To link a shared library named libopenblas.so, the flag -lopenblas is needed. To find the OpenBLAS headers, a -I/path/to/includedir is needed. And unless the library is installed in a directory that the linker searches by default, also -L and -Wl,-rpath flags are needed. For a source file test.c (e.g., the example code under Call CBLAS interface further down), the shared library can then be linked with:
gcc -o test test.c -I/your_path/OpenBLAS/include/ -L/your_path/OpenBLAS/lib -Wl,-rpath,/your_path/OpenBLAS/lib -lopenblas\nThe -Wl,-rpath,/your_path/OpenBLAS/lib linker flag can be omitted if you ran ldconfig to update linker cache, put /your_path/OpenBLAS/lib in /etc/ld.so.conf or a file in /etc/ld.so.conf.d, or installed OpenBLAS in a location that is part of the ld.so default search path (usually /lib, /usr/lib and /usr/local/lib). Alternatively, you can set the environment variable LD_LIBRARY_PATH to point to the folder that contains libopenblas.so. Otherwise, the build may succeed but at runtime loading the library will fail with a message like:
cannot open shared object file: no such file or directory\nMore flags may be needed, depending on how OpenBLAS was built:
libopenblas is multi-threaded, please add -lpthread.-lgfortran (other Fortran libraries may also be needed, e.g. -lquadmath). Note that if you only make calls to LAPACKE routines, i.e. your code has #include \"lapacke.h\" and makes calls to methods like LAPACKE_dgeqrf, then -lgfortran is not needed.Tip
Usually a pkg-config file (e.g., openblas.pc) is installed together with a libopenblas shared library. pkg-config is a tool that will tell you the exact flags needed for linking. For example:
$ pkg-config --cflags openblas\n-I/usr/local/include\n$ pkg-config --libs openblas\n-L/usr/local/lib -lopenblas\nLinking a static library is simpler - add the path to the static OpenBLAS library to the compile command:
gcc -o test test.c /your/path/libopenblas.a\nThis example shows calling cblas_dgemm in C:
#include <cblas.h>\n#include <stdio.h>\n\nvoid main()\n{\n int i=0;\n double A[6] = {1.0,2.0,1.0,-3.0,4.0,-1.0}; \n double B[6] = {1.0,2.0,1.0,-3.0,4.0,-1.0}; \n double C[9] = {.5,.5,.5,.5,.5,.5,.5,.5,.5}; \n cblas_dgemm(CblasColMajor, CblasNoTrans, CblasTrans,3,3,2,1,A, 3, B, 3,2,C,3);\n\n for(i=0; i<9; i++)\n printf(\"%lf \", C[i]);\n printf(\"\\n\");\n}\nTo compile this file, save it as test_cblas_dgemm.c and then run:
gcc -o test_cblas_open test_cblas_dgemm.c -I/your_path/OpenBLAS/include/ -L/your_path/OpenBLAS/lib -lopenblas -lpthread -lgfortran\ntest_cblas_open executable."},{"location":"user_manual/#call-blas-fortran-interface","title":"Call BLAS Fortran interface","text":"This example shows calling the dgemm Fortran interface in C:
#include \"stdio.h\"\n#include \"stdlib.h\"\n#include \"sys/time.h\"\n#include \"time.h\"\n\nextern void dgemm_(char*, char*, int*, int*,int*, double*, double*, int*, double*, int*, double*, double*, int*);\n\nint main(int argc, char* argv[])\n{\n int i;\n printf(\"test!\\n\");\n if(argc<4){\n printf(\"Input Error\\n\");\n return 1;\n }\n\n int m = atoi(argv[1]);\n int n = atoi(argv[2]);\n int k = atoi(argv[3]);\n int sizeofa = m * k;\n int sizeofb = k * n;\n int sizeofc = m * n;\n char ta = 'N';\n char tb = 'N';\n double alpha = 1.2;\n double beta = 0.001;\n\n struct timeval start,finish;\n double duration;\n\n double* A = (double*)malloc(sizeof(double) * sizeofa);\n double* B = (double*)malloc(sizeof(double) * sizeofb);\n double* C = (double*)malloc(sizeof(double) * sizeofc);\n\n srand((unsigned)time(NULL));\n\n for (i=0; i<sizeofa; i++)\n A[i] = i%3+1;//(rand()%100)/10.0;\n\n for (i=0; i<sizeofb; i++)\n B[i] = i%3+1;//(rand()%100)/10.0;\n\n for (i=0; i<sizeofc; i++)\n C[i] = i%3+1;//(rand()%100)/10.0;\n //#if 0\n printf(\"m=%d,n=%d,k=%d,alpha=%lf,beta=%lf,sizeofc=%d\\n\",m,n,k,alpha,beta,sizeofc);\n gettimeofday(&start, NULL);\n dgemm_(&ta, &tb, &m, &n, &k, &alpha, A, &m, B, &k, &beta, C, &m);\n gettimeofday(&finish, NULL);\n\n duration = ((double)(finish.tv_sec-start.tv_sec)*1000000 + (double)(finish.tv_usec-start.tv_usec)) / 1000000;\n double gflops = 2.0 * m *n*k;\n gflops = gflops/duration*1.0e-6;\n\n FILE *fp;\n fp = fopen(\"timeDGEMM.txt\", \"a\");\n fprintf(fp, \"%dx%dx%d\\t%lf s\\t%lf MFLOPS\\n\", m, n, k, duration, gflops);\n fclose(fp);\n\n free(A);\n free(B);\n free(C);\n return 0;\n}\nTo compile this file, save it as time_dgemm.c and then run:
gcc -o time_dgemm time_dgemm.c /your/path/libopenblas.a -lpthread\n./time_dgemm <m> <n> <k>, with m, n, and k input parameters to the time_dgemm executable. Note
When calling the Fortran interface from C, you have to deal with symbol name differences caused by compiler conventions. That is why the dgemm_ function call in the example above has a trailing underscore. This is what it looks like when using gcc/gfortran, however such details may change for different compilers. Hence it requires extra support code. The CBLAS interface may be more portable when writing C code.
When writing code that needs to be portable and work across different platforms and compilers, the above code example is not recommended for usage. Instead, we advise looking at how OpenBLAS (or BLAS in general, since this problem isn't specific to OpenBLAS) functions are called in widely used projects like Julia, SciPy, or R.
"},{"location":"user_manual/#troubleshooting","title":"Troubleshooting","text":"BIGNUMA=1.NO_AFFINITY=1 in Makefile.rule.make test is known to fail with a pthread_create error and an EAGAIN error code. However, it will be OK when you run the same testcase in a shell.OpenBLAS is an optimized Basic Linear Algebra Subprograms (BLAS) library based on GotoBLAS2 1.13 BSD version.
OpenBLAS implements low-level routines for performing linear algebra operations such as vector addition, scalar multiplication, dot products, linear combinations, and matrix multiplication. OpenBLAS makes these routines available on multiple platforms, covering server, desktop and mobile operating systems, as well as different architectures including x86, ARM, MIPS, PPC, RISC-V, and zarch.
The old GotoBLAS documentation can be found on GitHub.
"},{"location":"#license","title":"License","text":"OpenBLAS is licensed under the 3-clause BSD license. The full license can be found on GitHub.
"},{"location":"about/","title":"About","text":""},{"location":"about/#mailing-list","title":"Mailing list","text":"We have a GitHub discussions forum to discuss usage and development of OpenBLAS. We also have a Google group for users and a Google group for development of OpenBLAS.
"},{"location":"about/#acknowledgements","title":"Acknowledgements","text":"This work was or is partially supported by the following grants, contracts and institutions:
Over the course of OpenBLAS development, a number of donations were received. You can read OpenBLAS's statement of receipts and disbursement and cash balance in this Google doc (covers 2013-2016). A list of backers is available in BACKERS.md in the main repo.
"},{"location":"about/#donations","title":"Donations","text":"We welcome hardware donations, including the latest CPUs and motherboards.
"},{"location":"about/#open-source-users-of-openblas","title":"Open source users of OpenBLAS","text":"Prominent open source users of OpenBLAS include:
OpenBLAS is packaged in most major Linux distros, as well as general and numerical computing-focused packaging ecosystems like Nix, Homebrew, Spack and conda-forge.
OpenBLAS is used directly by libraries written in C, C++ and Fortran (and probably other languages), and directly by end users in those languages.
"},{"location":"about/#publications","title":"Publications","text":""},{"location":"about/#2013","title":"2013","text":"This page describes the Make-based build, which is the default/authoritative build method. Note that the OpenBLAS repository also supports building with CMake (not described here) - that generally works and is tested, however there may be small differences between the Make and CMake builds.
Warning
This page is made by someone who is not the developer and should not be considered as an official documentation of the build system. For getting the full picture, it is best to read the Makefiles and understand them yourself.
"},{"location":"build_system/#makefile-dep-graph","title":"Makefile dep graph","text":"Makefile \n| \n|----- Makefile.system # !!! this is included by many of the Makefiles in the subdirectories !!!\n| |\n| |===== Makefile.prebuild # This is triggered (not included) once by Makefile.system \n| | | # and runs before any of the actual library code is built.\n| | | # (builds and runs the \"getarch\" tool for cpu identification,\n| | | # runs the compiler detection scripts c_check and f_check) \n| | |\n| | ----- (Makefile.conf) [ either this or Makefile_kernel.conf is generated ] \n| | | { Makefile.system#L243 }\n| | ----- (Makefile_kernel.conf) [ temporary Makefile.conf during DYNAMIC_ARCH builds ]\n| |\n| |----- Makefile.rule # defaults for build options that can be given on the make command line\n| |\n| |----- Makefile.$(ARCH) # architecture-specific compiler options and OpenBLAS buffer size values\n|\n|~~~~~ exports/\n|\n|~~~~~ test/\n|\n|~~~~~ utest/ \n|\n|~~~~~ ctest/\n|\n|~~~~~ cpp_thread_test/\n|\n|~~~~~ kernel/\n|\n|~~~~~ ${SUBDIRS}\n|\n|~~~~~ ${BLASDIRS}\n|\n|~~~~~ ${NETLIB_LAPACK_DIR}{,/timing,/testing/{EIG,LIN}}\n|\n|~~~~~ relapack/\nMost of the tunable variables are found in Makefile.rule, along with their detailed descriptions. Most of the variables are detected automatically in Makefile.prebuild, if they are not set in the environment.
"},{"location":"build_system/#cpu-related","title":"CPU related","text":"ARCH - Target architecture (eg. x86_64)\nTARGET - Target CPU architecture, in case of DYNAMIC_ARCH=1 means library will not be usable on less capable CPUs\nTARGET_CORE - TARGET_CORE will override TARGET internally during each cpu-specific cycle of the build for DYNAMIC_ARCH\nDYNAMIC_ARCH - For building library for multiple TARGETs (does not lose any optimizations, but increases library size)\nDYNAMIC_LIST - optional user-provided subset of the DYNAMIC_CORE list in Makefile.system\nCC - TARGET C compiler used for compilation (can be cross-toolchains)\nFC - TARGET Fortran compiler used for compilation (can be cross-toolchains, set NOFORTRAN=1 if used cross-toolchain has no fortran compiler)\nAR, AS, LD, RANLIB - TARGET toolchain helpers used for compilation (can be cross-toolchains)\n\nHOSTCC - compiler of build machine, needed to create proper config files for target architecture\nHOST_CFLAGS - flags for build machine compiler\nBINARY - 32/64 bit library\n\nBUILD_SHARED - Create shared library\nBUILD_STATIC - Create static library\n\nQUAD_PRECISION - enable support for IEEE quad precision [ largely unimplemented leftover from GotoBLAS, do not use ]\nEXPRECISION - Obsolete option to use float80 of SSE on BSD-like systems\nINTERFACE64 - Build with 64bit integer representations to support large array index values [ incompatible with standard API ]\n\nBUILD_SINGLE - build the single-precision real functions of BLAS [and optionally LAPACK] \nBUILD_DOUBLE - build the double-precision real functions\nBUILD_COMPLEX - build the single-precision complex functions\nBUILD_COMPLEX16 - build the double-precision complex functions\n(all four types are included in the build by default when none was specifically selected)\n\nBUILD_BFLOAT16 - build the \"half precision brainfloat\" real functions \n\nUSE_THREAD - Use a multithreading backend (default to pthread)\nUSE_LOCKING - implement locking for thread safety even when USE_THREAD is not set (so that the singlethreaded library can\n safely be called from multithreaded programs)\nUSE_OPENMP - Use OpenMP as multithreading backend\nNUM_THREADS - define this to the maximum number of parallel threads you expect to need (defaults to the number of cores in the build cpu)\nNUM_PARALLEL - define this to the number of OpenMP instances that your code may use for parallel calls into OpenBLAS (default 1,see below)\nOpenBLAS uses a fixed set of memory buffers internally, used for communicating and compiling partial results from individual threads. For efficiency, the management array structure for these buffers is sized at build time - this makes it necessary to know in advance how many threads need to be supported on the target system(s).
With OpenMP, there is an additional level of complexity as there may be calls originating from a parallel region in the calling program. If OpenBLAS gets called from a single parallel region, it runs single-threaded automatically to avoid overloading the system by fanning out its own set of threads. In the case that an OpenMP program makes multiple calls from independent regions or instances in parallel, this default serialization is not sufficient as the additional caller(s) would compete for the original set of buffers already in use by the first call. So if multiple OpenMP runtimes call into OpenBLAS at the same time, then only one of them will be able to make progress while all the rest of them spin-wait for the one available buffer. Setting NUM_PARALLEL to the upper bound on the number of OpenMP runtimes that you can have in a process ensures that there are a sufficient number of buffer sets available.
OpenBLAS/ \n\u251c\u2500\u2500 benchmark Benchmark codes for BLAS\n\u251c\u2500\u2500 cmake CMakefiles\n\u251c\u2500\u2500 ctest Test codes for CBLAS interfaces\n\u251c\u2500\u2500 driver Implemented in C\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 level2\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 level3\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 mapper\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 others Memory management, threading, etc\n\u251c\u2500\u2500 exports Generate shared library\n\u251c\u2500\u2500 interface Implement BLAS and CBLAS interfaces (calling driver or kernel)\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 lapack\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 netlib\n\u251c\u2500\u2500 kernel Optimized assembly kernels for CPU architectures\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 alpha Original GotoBLAS kernels for DEC Alpha\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 arm ARMV5,V6,V7 kernels (including generic C codes used by other architectures)\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 arm64 ARMV8\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 generic General kernel codes written in plain C, parts used by many architectures.\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 ia64 Original GotoBLAS kernels for Intel Itanium\n\u2502 \u251c\u2500\u2500 mips\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 mips64\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 power\n| \u251c\u2500\u2500 riscv64\n| \u251c\u2500\u2500 simd Common code for Universal Intrinsics, used by some x86_64 and arm64 kernels\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 sparc\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 x86\n\u2502 \u251c\u2500\u2500 x86_64\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 zarch \n\u251c\u2500\u2500 lapack Optimized LAPACK codes (replacing those in regular LAPACK)\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 getf2\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 getrf\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 getrs\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 laswp\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 lauu2\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 lauum\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 potf2\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 potrf\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 trti2\n\u2502 \u251c\u2500\u2500 trtri\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 trtrs\n\u251c\u2500\u2500 lapack-netlib LAPACK codes from netlib reference implementation\n\u251c\u2500\u2500 reference BLAS Fortran reference implementation (unused)\n\u251c\u2500\u2500 relapack Elmar Peise's recursive LAPACK (implemented on top of regular LAPACK)\n\u251c\u2500\u2500 test Test codes for BLAS\n\u2514\u2500\u2500 utest Regression test\nA call tree for dgemm looks as follows:
interface/gemm.c\n \u2502\ndriver/level3/level3.c\n \u2502\ngemm assembly kernels at kernel/\nTo find the kernel currently used for a particular supported CPU, please check the corresponding kernel/$(ARCH)/KERNEL.$(CPU) file.
Here is an example for kernel/x86_64/KERNEL.HASWELL:
...\nDTRMMKERNEL = dtrmm_kernel_4x8_haswell.c\nDGEMMKERNEL = dgemm_kernel_4x8_haswell.S\n...\nKERNEL.HASWELL, OpenBLAS Haswell dgemm kernel file is dgemm_kernel_4x8_haswell.S."},{"location":"developers/#optimizing-gemm-for-a-given-hardware","title":"Optimizing GEMM for a given hardware","text":"Read the Goto paper to understand the algorithm
Goto, Kazushige; van de Geijn, Robert A. (2008). \"Anatomy of High-Performance Matrix Multiplication\". ACM Transactions on Mathematical Software 34 (3): Article 12
(The above link is available only to ACM members, but this and many related papers is also available on the pages of van de Geijn's FLAME project)
The driver/level3/level3.c is the implementation of Goto's algorithm. Meanwhile, you can look at kernel/generic/gemmkernel_2x2.c, which is a naive 2x2 register blocking gemm kernel in C. Then:
Mc, Kc, and Nc)Note that not all of the CPU-specific parameters in param.h are actively used in algorithms. DNUMOPT only appears as a scale factor in profiling output of the level3 syrk interface code, while its counterpart SNUMOPT (aliased as NUMOPT in common.h) is not used anywhere at all.
SYMV_P is only used in the generic kernels for the symv and chemv/zhemv functions - at least some of those are usually overridden by CPU-specific implementations, so if you start by cloning the existing implementation for a related CPU you need to check its KERNEL file to see if tuning SYMV_P would have any effect at all.
GEMV_UNROLL is only used by some older x86-64 kernels, so not all sections in param.h define it. Similarly, not all of the CPU parameters like L2 or L3 cache sizes are necessarily used in current kernels for a given model - by all indications the CPU identification code was imported from some other project originally.
We use tests for Netlib BLAS, CBLAS, and LAPACK. In addition, we use OpenBLAS-specific regression tests. They can be run with Make:
make -C test for BLAS testsmake -C ctest for CBLAS testsmake -C utest for OpenBLAS regression testsmake lapack-test for LAPACK testsWe also use the BLAS-Tester tests for regression testing. It is basically the ATLAS test suite adapted for building with OpenBLAS.
The project makes use of several Continuous Integration (CI) services conveniently interfaced with GitHub to automatically run tests on a number of platforms and build configurations.
Also note that the test suites included with \"numerically heavy\" projects like Julia, NumPy, SciPy, Octave or QuantumEspresso can be used for regression testing, when those projects are built such that they use OpenBLAS.
"},{"location":"developers/#benchmarking","title":"Benchmarking","text":"A number of benchmarking methods are used by OpenBLAS:
benchmark folder. They can be run locally through the Makefile in that directory. And the benchmark/scripts subdirectory contains similar benchmarks that use OpenBLAS via NumPy, SciPy, Octave and R.Benchmarking code for BLAS libraries, and specific performance analysis results, can be found in a number of places. For example:
Especially relevant for x86-64, a new CPU model may be a \"refresh\" (die shrink and/or different number of cores) within an existing model family without significant changes to its instruction set (e.g., Intel Skylake and Kaby Lake still are fundamentally the same architecture as Haswell, low end Goldmont etc. are Nehalem). In this case, compilation with the appropriate older TARGET will already lead to a satisfactory build.
To achieve autodetection of the new model, its CPUID (or an equivalent identifier) needs to be added in the cpuid_<architecture>.c relevant for its general architecture, with the returned name for the new type set appropriately. For x86, which has the most complex cpuid file, there are two functions that need to be edited: get_cpuname() to return, e.g., CPUTYPE_HASWELL and get_corename() for the (broader) core family returning, e.g., CORE_HASWELL.1
For architectures where DYNAMIC_ARCH builds are supported, a similar but simpler code section for the corresponding runtime detection of the CPU exists in driver/others/dynamic.c (for x86), and driver/others/dynamic_<arch>.c for other architectures. Note that for x86 the CPUID is compared after splitting it into its family, extended family, model and extended model parts, so the single decimal number returned by Linux in /proc/cpuinfo for the model has to be converted back to hexadecimal before splitting into its constituent digits. For example, 142 == 8E translates to extended model 8, model 14.
Usually it will be possible to start from an existing model, clone its KERNEL configuration file to the new name to use for this TARGET and eventually replace individual kernels with versions better suited for peculiarities of the new CPU model. In addition, it is necessary to add (or clone at first) the corresponding section of GEMM_UNROLL parameters in the top-level param.h, and possibly to add definitions such as USE_TRMM (governing whether TRMM functions use the respective GEMM kernel or a separate source file) to the Makefiles (and CMakeLists.txt) in the kernel directory. The new CPU name needs to be added to TargetList.txt, and the CPU auto-detection code used by the getarch helper program - contained in the cpuid_<architecture>.c file amended to include the CPUID (or equivalent) information processing required (see preceding section).
This endeavour is best started by cloning the entire support structure for 32-bit ARM, and within that the ARMv5 CPU in particular, as this is implemented through plain C kernels only. An example providing a convenient \"shopping list\" can be seen in pull request #1526.
This information ends up in the Makefile.conf and config.h files generated by getarch. Failure to set either will typically lead to a missing definition of the GEMM_UNROLL parameters later in the build, as getarch_2nd will be unable to find a matching parameter section in param.h.\u00a0\u21a9
Note
This document contains recommendations only - packagers and other redistributors are in charge of how OpenBLAS is built and distributed in their systems, and may have good reasons to deviate from the guidance given on this page. These recommendations are aimed at general packaging systems, with a user base that typically is large, open source (or freely available at least), and doesn't behave uniformly or that the packager is directly connected with.*
OpenBLAS has a large number of build-time options which can be used to change how it behaves at runtime, how artifacts or symbols are named, etc. Variation in build configuration can be necessary to acheive a given end goal within a distribution or as an end user. However, such variation can also make it more difficult to build on top of OpenBLAS and ship code or other packages in a way that works across many different distros. Here we provide guidance about the most important build options, what effects they may have when changed, and which ones to default to.
The Make and CMake build systems provide equivalent options and yield more or less the same artifacts, but not exactly (the CMake builds are still experimental). You can choose either one and the options will function in the same way, however the CMake outputs may require some renaming. To review available build options, see Makefile.rule or CMakeLists.txt in the root of the repository.
Build options typically fall into two categories: (a) options that affect the user interface, such as library and symbol names or APIs that are made available, and (b) options that affect performance and runtime behavior, such as threading behavior or CPU architecture-specific code paths. The user interface options are more important to keep aligned between distributions, while for the performance-related options there are typically more reasons to make choices that deviate from the defaults.
Here are recommendations for user interface related packaging choices where it is not likely to be a good idea to deviate (typically these are the default settings):
.pc) and CMake .cmake) dependency detection files. These files are used by build systems when users want to link against OpenBLAS, and there is no benefit of leaving them out.The LP64 (32-bit integer) interface is the default build, and has well-established C and Fortran APIs as determined by the reference (Netlib) BLAS and LAPACK libraries. The ILP64 (64-bit integer) interface however does not have a standard API: symbol names and shared/static library names can be produced in multiple ways, and this tends to make it difficult to use. As of today there is an agreed-upon way of choosing names for OpenBLAS between a number of key users/redistributors, which is the closest thing to a standard that there is now. However, there is an ongoing standardization effort in the reference BLAS and LAPACK libraries, which differs from the current OpenBLAS agreed-upon convention. In this section we'll aim to explain both.
Those two methods are fairly similar, and have a key thing in common: using a symbol suffix. This is good practice; it is recommended that if you distribute an ILP64 build, to have it use a symbol suffix containing 64 in the name. This avoids potential symbol clashes when different packages which depend on OpenBLAS load both an LP64 and an ILP64 library into memory at the same time.
This convention comprises the shared library name and the symbol suffix in the shared library. The symbol suffix to use is 64_, implying that the library name will be libopenblas64_.so and the symbols in that library end in 64_. The central issue where this was discussed is openblas#646, and adopters include Fedora, Julia, NumPy and SciPy - SuiteSparse already used it as well.
To build shared and static libraries with the currently recommended ILP64 conventions with Make:
$ make INTERFACE64=1 SYMBOLSUFFIX=64_\nThis will produce libraries named libopenblas64_.so|a, a pkg-config file named openblas64.pc, and CMake and header files.
Installing locally and inspecting the output will show a few more details:
$ make install PREFIX=$PWD/../openblas/make64 INTERFACE64=1 SYMBOLSUFFIX=64_\n$ tree . # output slightly edited down\n.\n\u251c\u2500\u2500 include\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 cblas.h\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 f77blas.h\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 lapacke_config.h\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 lapacke.h\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 lapacke_mangling.h\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 lapacke_utils.h\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 lapack.h\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 openblas_config.h\n\u2514\u2500\u2500 lib\n \u251c\u2500\u2500 cmake\n \u2502\u00a0\u00a0 \u2514\u2500\u2500 openblas\n \u2502\u00a0\u00a0 \u251c\u2500\u2500 OpenBLASConfig.cmake\n \u2502\u00a0\u00a0 \u2514\u2500\u2500 OpenBLASConfigVersion.cmake\n \u251c\u2500\u2500 libopenblas64_.a\n \u251c\u2500\u2500 libopenblas64_.so\n \u2514\u2500\u2500 pkgconfig\n \u2514\u2500\u2500 openblas64.pc\nA key point are the symbol names. These will equal the LP64 symbol names, then (for Fortran only) the compiler mangling, and then the 64_ symbol suffix. Hence to obtain the final symbol names, we need to take into account which Fortran compiler we are using. For the most common cases (e.g., gfortran, Intel Fortran, or Flang), that means appending a single underscore. In that case, the result is:
dgemm dgemm_64_ dgemm_64(...) dgemm_64_(...) cblas_dgemm cblas_dgemm64_ n/a cblas_dgemm64_(...) It is quite useful to have these symbol names be as uniform as possible across different packaging systems.
The equivalent build options with CMake are:
$ mkdir build && cd build\n$ cmake .. -DINTERFACE64=1 -DSYMBOLSUFFIX=64_ -DBUILD_SHARED_LIBS=ON -DBUILD_STATIC_LIBS=ON\n$ cmake --build . -j\nNote that the result is not 100% identical to the Make result. For example, the library name ends in _64 rather than 64_ - it is recommended to rename them to match the Make library names (also update the libsuffix entry in openblas64.pc to match that rename).
$ cmake --install . --prefix $PWD/../../openblas/cmake64\n$ tree .\n.\n\u251c\u2500\u2500 include\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 openblas64\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 cblas.h\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 f77blas.h\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 lapacke_config.h\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 lapacke_example_aux.h\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 lapacke.h\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 lapacke_mangling.h\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 lapacke_utils.h\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 lapack.h\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 openblas64\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u2514\u2500\u2500 lapacke_mangling.h\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 openblas_config.h\n\u2514\u2500\u2500 lib\n \u251c\u2500\u2500 cmake\n \u2502\u00a0\u00a0 \u2514\u2500\u2500 OpenBLAS64\n \u2502\u00a0\u00a0 \u251c\u2500\u2500 OpenBLAS64Config.cmake\n \u2502\u00a0\u00a0 \u251c\u2500\u2500 OpenBLAS64ConfigVersion.cmake\n \u2502\u00a0\u00a0 \u251c\u2500\u2500 OpenBLAS64Targets.cmake\n \u2502\u00a0\u00a0 \u2514\u2500\u2500 OpenBLAS64Targets-noconfig.cmake\n \u251c\u2500\u2500 libopenblas_64.a\n \u251c\u2500\u2500 libopenblas_64.so -> libopenblas_64.so.0\n \u2514\u2500\u2500 pkgconfig\n \u2514\u2500\u2500 openblas64.pc\nWhile the 64_ convention above got some adoption, it's slightly hacky and is implemented through the use of objcopy. An effort is ongoing for a more broadly adopted convention in the reference BLAS and LAPACK libraries, using (a) the _64 suffix, and (b) applying that suffix before rather than after Fortran compiler mangling. The central issue for this is lapack#666.
For the most common cases of compiler mangling (a single _ appended), the end result will be:
dgemm dgemm_64_ dgemm_64(...) dgemm_64_(...) cblas_dgemm cblas_dgemm_64 n/a cblas_dgemm_64(...) For other compiler mangling schemes, replace the trailing _ by the scheme in use.
The shared library name for this _64 convention should be libopenblas_64.so.
Note: it is not yet possible to produce an OpenBLAS build which employs this convention! Once reference BLAS and LAPACK with support for _64 have been released, a future OpenBLAS release will support it. For now, please use the older 64_ scheme and avoid using the name libopenblas_64.so; it should be considered reserved for future use of the _64 standard as prescribed by reference BLAS/LAPACK.
For these options there are multiple reasonable or common choices.
"},{"location":"distributing/#threading-related-options","title":"Threading related options","text":"OpenBLAS can be built as a multi-threaded or single-threaded library, with the default being multi-threaded. It's expected that the default libopenblas library is multi-threaded; if you'd like to also distribute single-threaded builds, consider naming them libopenblas_sequential.
OpenBLAS can be built with pthreads or OpenMP as the threading model, with the default being pthreads. Both options are commonly used, and the choice here should not influence the shared library name. The choice will be captured by the .pc file. E.g.,:
$ pkg-config --libs openblas\n-fopenmp -lopenblas\n\n$ cat openblas.pc\n...\nopenblas_config= ... USE_OPENMP=0 MAX_THREADS=24\nThe maximum number of threads users will be able to use is determined at build time by the NUM_THREADS build option. It defaults to 24, and there's a wide range of values that are reasonable to use (up to 256). 64 is a typical choice here; there is a memory footprint penalty that is linear in NUM_THREADS. Please see Makefile.rule for more details.
OpenBLAS contains a lot of CPU architecture-specific optimizations, hence when distributing to a user base with a variety of hardware, it is recommended to enable CPU architecture runtime detection. This will dynamically select optimized kernels for individual APIs. To do this, use the DYNAMIC_ARCH=1 build option. This is usually done on all common CPU families, except when there are known issues.
In case the CPU architecture is known (e.g. you're building binaries for macOS M1 users), it is possible to specify the target architecture directly with the TARGET= build option.
DYNAMIC_ARCH and TARGET are covered in more detail in the main README.md in this repository.
OpenBLAS is likely to be distributed in one of these distribution models:
The guidance on this page is most important for models (1) and (2). These links to build recipes for a representative selection of packaging systems may be helpful as a reference:
All major distributions do include LAPACK as of mid 2023 as far as we know. Older versions of Arch Linux did not, and that was known to cause problems.\u00a0\u21a9
OpenBLAS for the most part contains implementations of the reference (Netlib) BLAS, CBLAS, LAPACK and LAPACKE interfaces. A few OpenBLAS-specific functions are also provided however, which mostly can be seen as \"BLAS extensions\". This page documents those non-standard APIs.
"},{"location":"extensions/#blas-like-extensions","title":"BLAS-like extensions","text":"Routine Data Types Description ?axpby s,d,c,z likeaxpy with a multiplier for y ?gemm3m c,z gemm3m ?imatcopy s,d,c,z in-place transposition/copying ?omatcopy s,d,c,z out-of-place transposition/copying ?geadd s,d,c,z ATLAS-like matrix add B = α*A+β*B ?gemmt s,d,c,z gemm but only a triangular part updated"},{"location":"extensions/#bfloat16-functionality","title":"bfloat16 functionality","text":"BLAS-like and conversion functions for bfloat16 (available when OpenBLAS was compiled with BUILD_BFLOAT16=1):
void cblas_sbstobf16 converts a float array to an array of bfloat16 values by roundingvoid cblas_sbdtobf16 converts a double array to an array of bfloat16 values by roundingvoid cblas_sbf16tos converts a bfloat16 array to an array of floatsvoid cblas_dbf16tod converts a bfloat16 array to an array of doublesfloat cblas_sbdot computes the dot product of two bfloat16 arraysvoid cblas_sbgemv performs the matrix-vector operations of GEMV with the input matrix and X vector as bfloat16void cblas_sbgemm performs the matrix-matrix operations of GEMM with both input arrays containing bfloat16openblas_get_num_threadsopenblas_set_num_threadsint openblas_get_num_procs(void) returns the number of processors available on the system (may include \"hyperthreading cores\")int openblas_get_parallel(void) returns 0 for sequential use, 1 for platform-based threading and 2 for OpenMP-based threadingchar * openblas_get_config() returns the options OpenBLAS was built with, something like NO_LAPACKE DYNAMIC_ARCH NO_AFFINITY Haswellint openblas_set_affinity(int thread_index, size_t cpusetsize, cpu_set_t *cpuset) sets the CPU affinity mask of the given thread to the provided cpuset. Only available on Linux, with semantics identical to pthread_setaffinity_np.BLAS stands for Basic Linear Algebra Subprograms. BLAS provides standard interfaces for linear algebra, including BLAS1 (vector-vector operations), BLAS2 (matrix-vector operations), and BLAS3 (matrix-matrix operations). In general, BLAS is the computational kernel (\"the bottom of the food chain\") in linear algebra or scientific applications. Thus, if BLAS implementation is highly optimized, the whole application can get substantial benefit.
"},{"location":"faq/#what-functions-are-there-and-how-can-i-call-them-from-my-c-code","title":"What functions are there and how can I call them from my C code?","text":"As BLAS is a standardized interface, you can refer to the documentation of its reference implementation at netlib.org. Calls from C go through its CBLAS interface, so your code will need to include the provided cblas.h in addition to linking with -lopenblas. A single-precision matrix multiplication will look like
#include <cblas.h>\n...\ncblas_sgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, M, N, K, 1.0, A, K, B, N, 0.0, result, N);\nOpenBLAS is an open source BLAS library forked from the GotoBLAS2-1.13 BSD version. Since Mr. Kazushige Goto left TACC, GotoBLAS is no longer being maintained. Thus, we created this project to continue developing OpenBLAS/GotoBLAS.
"},{"location":"faq/#whats-the-difference-between-openblas-and-gotoblas","title":"What's the difference between OpenBLAS and GotoBLAS?","text":"In OpenBLAS 0.2.0, we optimized level 3 BLAS on the Intel Sandy Bridge 64-bit OS. We obtained a performance comparable with that Intel MKL.
We optimized level 3 BLAS performance on the ICT Loongson-3A CPU. It outperformed GotoBLAS by 135% in a single thread and 120% in 4 threads.
We fixed some GotoBLAS bugs including a SEGFAULT bug on the new Linux kernel, MingW32/64 bugs, and a ztrmm computing error bug on Intel Nehalem.
We also added some minor features, e.g. supporting \"make install\", compiling without LAPACK and upgrading the LAPACK version to 3.4.2.
You can find the full list of modifications in Changelog.txt.
"},{"location":"faq/#where-do-parameters-gemm_p-gemm_q-gemm_r-come-from","title":"Where do parameters GEMM_P, GEMM_Q, GEMM_R come from?","text":"The detailed explanation is probably in the original publication authored by Kazushige Goto - Goto, Kazushige; van de Geijn, Robert A; Anatomy of high-performance matrix multiplication. ACM Transactions on Mathematical Software (TOMS). Volume 34 Issue 3, May 2008 While this article is paywalled and too old for preprints to be available on arxiv.org, more recent publications like https://arxiv.org/pdf/1609.00076 contain at least a brief description of the algorithm. In practice, the values are derived by experimentation to yield the block sizes that give the highest performance. A general rule of thumb for selecting a starting point seems to be that PxQ is about half the size of L2 cache.
"},{"location":"faq/#how-can-i-report-a-bug","title":"How can I report a bug?","text":"Please file an issue at this issue page or send mail to the OpenBLAS mailing list.
Please provide the following information: CPU, OS, compiler, OpenBLAS version and any compiling flags you used (Makefile.rule). In addition, please describe how to reproduce this bug.
"},{"location":"faq/#how-to-reference-openblas","title":"How to reference OpenBLAS.","text":"You can reference our papers in this page. Alternatively, you can cite the OpenBLAS homepage http://www.openblas.net.
"},{"location":"faq/#how-can-i-use-openblas-in-multi-threaded-applications","title":"How can I use OpenBLAS in multi-threaded applications?","text":"If your application is already multi-threaded, it will conflict with OpenBLAS multi-threading. Thus, you must set OpenBLAS to use single thread as following.
If the application is parallelized by OpenMP, please build OpenBLAS with USE_OPENMP=1
With the increased availability of fast multicore hardware it has unfortunately become clear that the thread management provided by OpenMP is not sufficient to prevent race conditions when OpenBLAS was built single-threaded by USE_THREAD=0 and there are concurrent calls from multiple threads to OpenBLAS functions. In this case, it is vital to also specify USE_LOCKING=1 (introduced with OpenBLAS 0.3.7).
"},{"location":"faq/#does-openblas-support-sparse-matrices-andor-vectors","title":"Does OpenBLAS support sparse matrices and/or vectors ?","text":"OpenBLAS implements only the standard (dense) BLAS and LAPACK functions with a select few extensions popularized by Intel's MKL. Some cases can probably be made to work using e.g. GEMV or AXPBY, in general using a dedicated package like SuiteSparse (which can make use of OpenBLAS or equivalent for standard operations) is recommended.
"},{"location":"faq/#what-support-is-there-for-recent-pc-hardware-what-about-gpu","title":"What support is there for recent PC hardware ? What about GPU ?","text":"As OpenBLAS is a volunteer project, it can take some time for the combination of a capable developer, free time, and particular hardware to come along, even for relatively common processors. Starting from 0.3.1, support is being added for AVX 512 (TARGET=SKYLAKEX), requiring a compiler that is capable of handling avx512 intrinsics. While AMD Zen processors should be autodetected by the build system, as of 0.3.2 they are still handled exactly like Intel Haswell. There once was an effort to build an OpenCL implementation that one can still find at https://github.com/xianyi/clOpenBLAS , but work on this stopped in 2015.
"},{"location":"faq/#how-about-the-level-3-blas-performance-on-intel-sandy-bridge","title":"How about the level 3 BLAS performance on Intel Sandy Bridge?","text":"We obtained a performance comparable with Intel MKL that actually outperformed Intel MKL in some cases. Here is the result of the DGEMM subroutine's performance on Intel Core i5-2500K Windows 7 SP1 64-bit:
"},{"location":"faq/#os-and-compiler","title":"OS and Compiler","text":""},{"location":"faq/#how-can-i-call-an-openblas-function-in-microsoft-visual-studio","title":"How can I call an OpenBLAS function in Microsoft Visual Studio?","text":"Please read this page.
"},{"location":"faq/#how-can-i-use-cblas-and-lapacke-without-c99-complex-number-support-eg-in-visual-studio","title":"How can I use CBLAS and LAPACKE without C99 complex number support (e.g. in Visual Studio)?","text":"Zaheer has fixed this bug. You can now use the structure instead of C99 complex numbers. Please read this issue page for details.
This issue is for using LAPACKE in Visual Studio.
"},{"location":"faq/#i-get-a-segfault-with-multi-threading-on-linux-whats-wrong","title":"I get a SEGFAULT with multi-threading on Linux. What's wrong?","text":"This may be related to a bug in the Linux kernel 2.6.32 (?). Try applying the patch segaults.patch to disable mbind using
patch < segfaults.patch\nand see if the crashes persist. Note that this patch will lead to many compiler warnings.
"},{"location":"faq/#when-i-make-the-library-there-is-no-such-instruction-xgetbv-error-whats-wrong","title":"When I make the library, there is no such instruction: `xgetbv' error. What's wrong?","text":"Please use GCC 4.4 and later version. This version supports xgetbv instruction. If you use the library for Sandy Bridge with AVX instructions, you should use GCC 4.6 and later version.
On Mac OS X, please use Clang 3.1 and later version. For example, make CC=clang
For the compatibility with old compilers (GCC < 4.4), you can enable NO_AVX flag. For example, make NO_AVX=1
"},{"location":"faq/#my-build-fails-due-to-the-linker-error-multiple-definition-of-dlamc3_-what-is-the-problem","title":"My build fails due to the linker error \"multiple definition of `dlamc3_'\". What is the problem?","text":"This linker error occurs if GNU patch is missing or if our patch for LAPACK fails to apply.
Background: OpenBLAS implements optimized versions of some LAPACK functions, so we need to disable the reference versions. If this process fails we end with duplicated implementations of the same function.
"},{"location":"faq/#my-build-worked-fine-and-passed-all-tests-but-running-make-lapack-test-ends-with-segfaults","title":"My build worked fine and passed all tests, but runningmake lapack-test ends with segfaults","text":"Some of the LAPACK tests, notably in xeigtstz, try to allocate around 10MB on the stack. You may need to use ulimit -s to change the default limits on your system to allow this.
make lapack-test ends with a number of errors in the summary report","text":"The LAPACK tests were primarily created to test the validity of the Reference-LAPACK implementation, which is implemented in unoptimized, single-threaded Fortran code. This makes it very sensitive to small numerical deviations that can result from the use of specialized cpu instructions that combine multiplications and additions without intermediate rounding and storing to memory (FMA), or from changing the order of mathematical operations by splitting an original problem workload into smaller tasks that are solved in parallel. As a result, you may encounter a small number of errors in the \"numerical\" column of the summary table at the end of the make lapack-test run - this is usually nothing to worry about, and the exact number and distribution of errors among the four data types will often vary with the optimization flags you supplied to the compiler, or the cpu model for which you built OpenBLAS. Sporadic errors in the column labeled other are normally the sign of failed convergence of iterative diagonalizations for the same reasons just mentioned. A more detailed error report is stored in the file testing_results.txt - this should be consulted in case of doubt. Care should be taken if you encounter numerical errors in the hundreds, or other errors accompanied by the LAPACK error message \"on entry to function_name parameter X had an illegal value\" that signals a problem with argument passing between individual functions. (See also this issue in the issue tracker on github for additional discussion, examples and links)
You can define the OPENBLAS_MAIN_FREE or GOTOBLAS_MAIN_FREE environment variable to disable threading affinity on runtime. For example, before the running,
export OPENBLAS_MAIN_FREE=1\nAlternatively, you can disable affinity feature with enabling NO_AFFINITY=1 in Makefile.rule.
"},{"location":"faq/#how-to-solve-undefined-reference-errors-when-statically-linking-against-libopenblasa","title":"How to solve undefined reference errors when statically linking against libopenblas.a","text":"On Linux, if OpenBLAS was compiled with threading support (USE_THREAD=1 by default), custom programs statically linked against libopenblas.a should also link to the pthread library e.g.:
gcc -static -I/opt/OpenBLAS/include -L/opt/OpenBLAS/lib -o my_program my_program.c -lopenblas -lpthread\nFailing to add the -lpthread flag will cause errors such as:
/opt/OpenBLAS/libopenblas.a(memory.o): In function `_touch_memory':\nmemory.c:(.text+0x15): undefined reference to `pthread_mutex_lock'\nmemory.c:(.text+0x41): undefined reference to `pthread_mutex_unlock'\n/opt/OpenBLAS/libopenblas.a(memory.o): In function `openblas_fork_handler':\nmemory.c:(.text+0x440): undefined reference to `pthread_atfork'\n/opt/OpenBLAS/libopenblas.a(memory.o): In function `blas_memory_alloc':\nmemory.c:(.text+0x7a5): undefined reference to `pthread_mutex_lock'\nmemory.c:(.text+0x825): undefined reference to `pthread_mutex_unlock'\n/opt/OpenBLAS/libopenblas.a(memory.o): In function `blas_shutdown':\nmemory.c:(.text+0x9e1): undefined reference to `pthread_mutex_lock'\nmemory.c:(.text+0xa6e): undefined reference to `pthread_mutex_unlock'\n/opt/OpenBLAS/libopenblas.a(blas_server.o): In function `blas_thread_server':\nblas_server.c:(.text+0x273): undefined reference to `pthread_mutex_lock'\nblas_server.c:(.text+0x287): undefined reference to `pthread_mutex_unlock'\nblas_server.c:(.text+0x33f): undefined reference to `pthread_cond_wait'\n/opt/OpenBLAS/libopenblas.a(blas_server.o): In function `blas_thread_init':\nblas_server.c:(.text+0x416): undefined reference to `pthread_mutex_lock'\nblas_server.c:(.text+0x4be): undefined reference to `pthread_mutex_init'\nblas_server.c:(.text+0x4ca): undefined reference to `pthread_cond_init'\nblas_server.c:(.text+0x4e0): undefined reference to `pthread_create'\nblas_server.c:(.text+0x50f): undefined reference to `pthread_mutex_unlock'\n...\nThe -lpthread is not required when linking dynamically against libopenblas.so.0.
Minimum requirement to actually run AVX2-enabled software like OpenBLAS is kernel-2.6.32-358, shipped with EL6U4 in 2013
The binutils package from RHEL6 does not know the instruction vpermpd or any other AVX2 instruction. You can download a newer binutils package from Enterprise Linux software collections, following instructions here: https://www.softwarecollections.org/en/scls/rhscl/devtoolset-3/ After configuring repository you need to install devtoolset-?-binutils to get later usable binutils package
$ yum search devtoolset-\\?-binutils\n$ sudo yum install devtoolset-3-binutils\n$ scl --list\ndevtoolset-3\nrh-python35\n$ scl enable devtoolset-3 -- make DYNAMIC_ARCH=1\nBy default, QEMU reports the CPU as \"QEMU Virtual CPU version 2.2.0\", which shares CPUID with existing 32bit CPU even in 64bit virtual machine, and OpenBLAS recognizes it as PENTIUM2. Depending on the exact combination of CPU features the hypervisor choses to expose, this may not correspond to any CPU that exists, and OpenBLAS will error when trying to build. To fix this, pass -cpu host or -cpu passthough to QEMU, or another CPU model. Similarly, the XEN hypervisor may not pass through all features of the host cpu while reporting the cpu type itself correctly, which can lead to compiler error messages about an \"ABI change\" when compiling AVX512 code. Again changing the Xen configuration by running e.g. \"xen-cmdline --set-xen cpuid=avx512\" should get around this (as would building OpenBLAS for an older cpu lacking that particular feature, e.g. TARGET=HASWELL)
Trying to compile OpenBLAS with IBM XL ends with error messages about unknown register names\nlike \"vs32\". Working around these by using known alternate names for the vector registers only leads to another assembler error about unsupported constraints. This is a known deficiency in the IBM compiler at least up to and including 16.1.0 (and in the POWER version of clang, from which it is derived) - use gcc instead. (See issues #1078 and #1699 for related discussions)
"},{"location":"faq/#replacing-system-blasupdating-apt-openblas-in-mintubuntudebian","title":"Replacing system BLAS/updating APT OpenBLAS in Mint/Ubuntu/Debian","text":"Debian and Ubuntu LTS versions provide OpenBLAS package which is not updated after initial release, and under circumstances one might want to use more recent version of OpenBLAS e.g. to get support for newer CPUs
Ubuntu and Debian provides 'alternatives' mechanism to comfortably replace BLAS and LAPACK libraries systemwide.
After successful build of OpenBLAS (with DYNAMIC_ARCH set to 1)
$ make clean\n$ make DYNAMIC_ARCH=1\n$ sudo make DYNAMIC_ARCH=1 install\n$ sudo apt install libblas-dev liblapack-dev\n$ sudo update-alternatives --install /usr/lib/libblas.so.3 libblas.so.3 /opt/OpenBLAS/lib/libopenblas.so.0 41 \\\n --slave /usr/lib/liblapack.so.3 liblapack.so.3 /opt/OpenBLAS/lib/libopenblas.so.0\n$ sudo update-alternatives --remove libblas.so.3 /opt/OpenBLAS/lib/libopenblas.so.0\n$ update-alternatives --display libblas.so.3 to find out what layout your system has."},{"location":"faq/#i-built-openblas-for-use-with-some-other-software-but-that-software-cannot-find-it","title":"I built OpenBLAS for use with some other software, but that software cannot find it","text":"Openblas installs as a single library named libopenblas.so, while some programs may be searching for a separate libblas.so and liblapack.so so you may need to create appropriate symbolic links (ln -s libopenblas.so libblas.so; ln -s libopenblas.so liblapack.so) or copies. Also make sure that the installation location (usually /opt/OpenBLAS/lib or /usr/local/lib) is among the library search paths of your system.
You probably tried to include a cblas.h that you simply copied from the OpenBLAS source, instead you need to run make install after building OpenBLAS and then use the modified cblas.h that this step builds in the installation path (usually either /usr/local/include, /opt/OpenBLAS/include or whatever you specified as PREFIX= on the make install)
This is due to different interpretations of the (informal) standard for passing characters as arguments between C and FORTRAN functions. As the method for storing text differs in the two languages, when C calls Fortran the text length is passed as an \"invisible\" additional parameter. Historically, this has not been required when the text is just a single character, so older code like the Reference-LAPACK bundled with OpenBLAS does not do it. Recently gcc's checking has changed to require it, but there is no consensus yet if and how the existing LAPACK (and many other codebases) should adapt. (And for actual compilation, gcc has mostly backtracked and provided compatibility options - hence the default build settings in the OpenBLAS Makefiles add -fno-optimize-sibling-calls to the gfortran options to prevent miscompilation with \"affected\" versions. See ticket 2154 in the issue tracker for more details and links)
"},{"location":"faq/#build-fails-with-lots-of-errors-about-undefined-gemm_unroll_m","title":"Build fails with lots of errors about undefined ?GEMM_UNROLL_M","text":"Your cpu is apparently too new to be recognized by the build scripts, so they failed to assign appropriate parameters for the block algorithm. Do a make clean and try again with TARGET set to one of the cpu models listed in TargetList.txt - for x86_64 this will usually be HASWELL.
This is a limitation in the maximum length of a command on OSX, coupled with how CMAKE works. You should be able to work around this by adding the option -DCMAKE_Fortran_USE_RESPONSE_FILE_FOR_OBJECTS=1 to your CMAKE arguments.
There have been a few reports of wrong calculation results and build-time test failures when building in a container environment managed by the OSX version of Docker Desktop, which uses the xhyve virtualizer underneath. Judging from these reports, AVX2 support in xhyve appears to be subtly broken but a corresponding ticket in the xhyve issue tracker has not drawn any reaction or comment since 2019. Therefore it is strongly recommended to build OpenBLAS with the NO_AVX2=1 option when inside a container under (or for later use with) the Docker Desktop environment on Intel-based Apple hardware.
"},{"location":"faq/#usage","title":"Usage","text":""},{"location":"faq/#program-is-terminated-because-you-tried-to-allocate-too-many-memory-regions","title":"Program is Terminated. Because you tried to allocate too many memory regions","text":"In OpenBLAS, we mange a pool of memory buffers and allocate the number of buffers as the following.
#define NUM_BUFFERS (MAX_CPU_NUMBER * 2)\nPlease build OpenBLAS with larger NUM_THREADS. For example, make NUM_THREADS=32 or make NUM_THREADS=64. In Makefile.system, we will set MAX_CPU_NUMBER=NUM_THREADS.
The environment variable which control the kernel selection is OPENBLAS_CORETYPE (see driver/others/dynamic.c) e.g. export OPENBLAS_CORETYPE=Haswell. And the function char* openblas_get_corename() returns the used target.
This symbol gets defined only when OpenBLAS is built with \"make DYNAMIC_ARCH=1\" (which is what distributors will choose to ensure support for more than just one CPU type).
"},{"location":"faq/#how-can-i-find-out-at-runtime-what-options-the-library-was-built-with","title":"How can I find out at runtime what options the library was built with ?","text":"OpenBLAS has two utility functions that may come in here:
openblas_get_parallel() will return 0 for a single-threaded library, 1 if multithreading without OpenMP, 2 if built with USE_OPENMP=1
openblas_get_config() will return a string containing settings such as USE64BITINT or DYNAMIC_ARCH that were active at build time, as well as the target cpu (or in case of a dynamic_arch build, the currently detected one).
"},{"location":"faq/#after-making-openblas-i-find-that-the-static-library-is-multithreaded-but-the-dynamic-one-is-not","title":"After making OpenBLAS, I find that the static library is multithreaded, but the dynamic one is not ?","text":"The shared OpenBLAS library you built is probably working fine as well, but your program may be picking up a different (probably single-threaded) version from one of the standard system paths like /usr/lib on startup. Running ldd /path/to/your/program will tell you which library the linkage loader will actually use.
Specifying the \"correct\" library location with the -L flag (like -L /opt/OpenBLAS/lib) when linking your program only defines which library will be used to see if all symbols can be resolved, you will need to add an rpath entry to the binary (using -Wl,rpath=/opt/OpenBLAS/lib) to make it request searching that location. Alternatively, remove the \"wrong old\" library (if you can), or set LD_LIBRARY_PATH to the desired location before running your program.
You need to edit file src/cuda/cuda_dgemm.c in the NVIDIA version of HPL, change the \"handle2\" and \"handle\" dlopen calls to use libopenblas.so instead of libmkl_intel_lp64.so, and add an trailing underscore in the dlsym lines for dgemm_mkl and dtrsm_mkl (like dgemm_mkl = (void(*)())dlsym(handle, \u201cdgemm_\u201d);)
The power saving mechanisms of your board may have shut down some cores, making them invisible to OpenBLAS in its startup phase. Try bringing them online before starting your calculation.
"},{"location":"faq/#speed-varies-wildly-between-individual-runs-on-a-typical-armv8-smartphone-processor","title":"Speed varies wildly between individual runs on a typical ARMV8 smartphone processor","text":"Check the technical specifications, it could be that the SoC combines fast and slow cpus and threads can end up on either. In that case, binding the process to specific cores e.g. by setting OMP_PLACES=cores may help. (You may need to experiment with OpenMP options, it has been reported that using OMP_NUM_THREADS=2 OMP_PLACES=cores caused a huge drop in performance on a 4+4 core chip while OMP_NUM_THREADS=2 OMP_PLACES=cores(2) worked as intended - as did OMP_PLACES=cores with 4 threads)
Multithreading support in OpenBLAS requires the use of internal buffers for sharing partial results, the number and size of which is defined at compile time. Unless you specify NUM_THREADS in your make or cmake command, the build scripts try to autodetect the number of cores available in your build host to size the library to match. This unfortunately means that if you move the resulting binary from a small \"front-end node\" to a larger \"compute node\" later, it will still be limited to the hardware capabilities of the original system. The solution is to set NUM_THREADS to a number big enough to encompass the biggest systems you expect to run the binary on - at runtime, it will scale down the maximum number of threads it uses to match the number of cores physically available.
"},{"location":"faq/#getting-elf-load-command-addressoffset-not-properly-aligned-when-loading-libopenblasso","title":"Getting \"ELF load command address/offset not properly aligned\" when loading libopenblas.so","text":"If you get a message \"error while loading shared libraries: libopenblas.so.0: ELF load command address/offset not properly aligned\" when starting a program that is (dynamically) linked to OpenBLAS, this is very likely due to a bug in the GNU linker (ld) that is part of the GNU binutils package. This error was specifically observed on older versions of Ubuntu Linux updated with the (at the time) most recent binutils version 2.38, but an internet search turned up sporadic reports involving various other libraries dating back several years. A bugfix was created by the binutils developers and should be available in later versions of binutils.(See issue 3708 for details)
"},{"location":"faq/#using-openblas-with-openmp","title":"Using OpenBLAS with OpenMP","text":"OpenMP provides its own locking mechanisms, so when your code makes BLAS/LAPACK calls from inside OpenMP parallel regions it is imperative that you use an OpenBLAS that is built with USE_OPENMP=1, as otherwise deadlocks might occur. Furthermore, OpenBLAS will automatically restrict itself to using only a single thread when called from an OpenMP parallel region. When it is certain that calls will only occur from the main thread of your program (i.e. outside of omp parallel constructs), a standard pthreads build of OpenBLAS can be used as well. In that case it may be useful to tune the linger behaviour of idle threads in both your OpenMP program (e.g. set OMP_WAIT_POLICY=passive) and OpenBLAS (by redefining the THREAD_TIMEOUT variable at build time, or setting the environment variable OPENBLAS_THREAD_TIMEOUT smaller than the default 26) so that the two alternating thread pools do not unnecessarily hog the cpu during the handover.
"},{"location":"install/","title":"Install OpenBLAS","text":"OpenBLAS can be installed through package managers or from source. If you only want to use OpenBLAS rather than make changes to it, we recommend installing a pre-built binary package with your package manager of choice.
This page contains an overview of installing with package managers as well as from source. For the latter, see further down on this page.
"},{"location":"install/#installing-with-a-package-manager","title":"Installing with a package manager","text":"Note
Almost every package manager provides OpenBLAS packages; the list on this page is not comprehensive. If your package manager of choice isn't shown here, please search its package database for openblas or libopenblas.
On Linux, OpenBLAS can be installed with the system package manager, or with a package manager like Conda (or alternative package managers for the conda-forge ecosystem, like Mamba, Micromamba, or Pixi), Spack, or Nix. For the latter set of tools, the package name in all cases is openblas. Since package management in quite a few of these tools is declarative (i.e., managed by adding openblas to a metadata file describing the dependencies for your project or environment), we won't attempt to give detailed instructions for these tools here.
Linux distributions typically split OpenBLAS up in two packages: one containing the library itself (typically named openblas or libopenblas), and one containing headers, pkg-config and CMake files (typically named the same as the package for the library with -dev or -devel appended; e.g., openblas-devel). Please keep in mind that if you want to install OpenBLAS in order to use it directly in your own project, you will need to install both of those packages.
Distro-specific installation commands:
Debian/Ubuntu/Mint/KaliopenSUSE/SLEFedora/CentOS/RHELArch/Manjaro/Antergos$ sudo apt update\n$ sudo apt install libopenblas-dev\nupdate-alternatives mechanism: $ sudo update-alternatives --config libblas.so.3\n$ sudo zypper refresh\n$ sudo zypper install openblas-devel\nOpenBLAS can be configured as the default BLAS through the update-alternatives mechanism:
$ sudo update-alternatives --config libblas.so.3\n$ dnf check-update\n$ dnf install openblas-devel\nWarning
Fedora does not ship the pkg-config files for OpenBLAS. Instead, it wants you to link against FlexiBLAS (which uses OpenBLAS by default as its backend on Fedora), which you can install with:
$ dnf install flexiblas-devel\nFor CentOS and RHEL, OpenBLAS packages are provided via the Fedora EPEL repository. After adding that repository and its repository keys, you can install openblas-devel with either dnf or yum.
$ sudo pacman -S openblas\nOpenBLAS can be installed with conda (or mamba, micromamba, or pixi) from conda-forge:
conda install openblas\nConda-forge provides a method for switching the default BLAS implementation used by all packages. To use that for OpenBLAS, install libblas=*=*openblas (see the docs on this mechanism for more details).
OpenBLAS can be installed with vcpkg:
# In classic mode:\nvcpkg install openblas\n\n# Or in manifest mode:\nvcpkg add port openblas\nWindows is the only platform for which binaries are made available by the OpenBLAS project itself. They can be downloaded from the GitHub Releases](https://github.com/OpenMathLib/OpenBLAS/releases) page. These binaries are built with MinGW, using the following build options:
NUM_THREADS=64 TARGET=GENERIC DYNAMIC_ARCH=1 DYNAMIC_OLDER=1 CONSISTENT_FPCSR=1 INTERFACE=0\n.zip bundle there (note that you will need to edit the provided openblas.pc and OpenBLASConfig.cmake to reflect the installation path on your computer, as distributed they have \"win\" or \"win64\" reflecting the local paths on the system they were built on). Note that the same binaries can be downloaded from SourceForge; this is mostly of historical interest.
"},{"location":"install/#macos","title":"macOS","text":"To install OpenBLAS with a package manager on macOS, run:
HomebrewMacPortsConda-forge% brew install openblas\n% sudo port install OpenBLAS-devel\n% conda install openblas\nConda-forge provides a method for switching the default BLAS implementation used by all packages. To use that for OpenBLAS, install libblas=*=*openblas (see the docs on this mechanism for more details).
You can install OpenBLAS from the FreeBSD Ports collection:
pkg install openblas\nWe recommend download the latest stable version from the GitHub Releases page, or checking it out from a git tag, rather than a dev version from the develop branch.
Tip
The User manual contains a section with detailed information on compiling OpenBLAS, including how to customize builds and how to cross-compile. Please read that documentation first. This page contains only platform-specific build information, and assumes you already understand the general build system invocations to build OpenBLAS, with the specific build options you want to control multi-threading and other non-platform-specific behavior).
"},{"location":"install/#linux-and-macos","title":"Linux and macOS","text":"Ensure you have C and Fortran compilers installed, then simply type make to compile the library. There are no other build dependencies, nor unusual platform-specific environment variables to set or other system setup to do.
Note
When building in an emulator (KVM, QEMU, etc.), please make sure that the combination of CPU features exposed to the virtual environment matches that of an existing CPU to allow detection of the CPU model to succeed. (With qemu, this can be done by passing -cpu host or a supported model name at invocation).
We support building OpenBLAS with either MinGW or Visual Studio on Windows. Using MSVC will yield an OpenBLAS build with the Windows platform-native ABI. Using MinGW will yield a different ABI. We'll describe both methods in detail in this section, since the process for each is quite different.
"},{"location":"install/#visual-studio-native-windows-abi","title":"Visual Studio & native Windows ABI","text":"For Visual Studio, you can use CMake to generate Visual Studio solution files; note that you will need at least CMake 3.11 for linking to work correctly).
Note that you need a Fortran compiler if you plan to build and use the LAPACK functions included with OpenBLAS. The sections below describe using either flang as an add-on to clang/LLVM or gfortran as part of MinGW for this purpose. If you want to use the Intel Fortran compiler (ifort or ifx) for this, be sure to also use the Intel C compiler (icc or icx) for building the C parts, as the ABI imposed by ifort is incompatible with MSVC
A fully-optimized OpenBLAS that can be statically or dynamically linked to your application can currently be built for the 64-bit architecture with the LLVM compiler infrastructure. We're going to use Miniconda3 to grab all of the tools we need, since some of them are in an experimental status. Before you begin, you'll need to have Microsoft Visual Studio 2015 or newer installed.
winget install --id Anaconda.Miniconda3, or easily download from conda.io.%USERPROFILE%\\miniconda3\\shell\\condabin\\conda-hook.ps1.cd to change to the directory where you want to build OpenBLAS.conda update -n base conda\nconda config --add channels conda-forge\nconda install -y cmake flang clangdev perl libflang ninja\nStill in the Anaconda Command Prompt window, activate the 64-bit MSVC environment with vcvarsall x64. On Windows 11 with Visual Studio 2022, this would be done by invoking:
\"c:\\Program Files\\Microsoft Visual Studio\\2022\\Community\\vc\\Auxiliary\\Build\\vcvars64.bat\"\nWith VS2019, the command should be the same (except for the year number of course). For other versions of MSVC, please check the Visual Studio documentation for exactly how to invoke the vcvars64.bat script.
Confirm that the environment is active by typing link. This should return a long list of possible options for the link command. If it just returns \"command not found\" or similar, review and retype the call to vcvars64.bat.
Note
if you are working from a Visual Studio command prompt window instead (so that you do not have to do the vcvars call), you need to invoke conda activate so that CONDA_PREFIX etc. get set up correctly before proceeding to step 6. Failing to do so will lead to link errors like libflangmain.lib not getting found later in the build.
Now configure the project with CMake. Starting in the project directory, execute the following:
set \"LIB=%CONDA_PREFIX%\\Library\\lib;%LIB%\"\nset \"CPATH=%CONDA_PREFIX%\\Library\\include;%CPATH%\"\nmkdir build\ncd build\ncmake .. -G \"Ninja\" -DCMAKE_CXX_COMPILER=clang-cl -DCMAKE_C_COMPILER=clang-cl -DCMAKE_Fortran_COMPILER=flang -DCMAKE_MT=mt -DBUILD_WITHOUT_LAPACK=no -DNOFORTRAN=0 -DDYNAMIC_ARCH=ON -DCMAKE_BUILD_TYPE=Release\nYou may want to add further options in the cmake command here. For instance, the default only produces a static .lib version of the library. If you would rather have a DLL, add -DBUILD_SHARED_LIBS=ON above. Note that this step only creates some command files and directories, the actual build happens next.
Build the project:
cmake --build . --config Release\nlib directory, and various build-time tests in the test, ctest and openblas_utest directories. However it will not separate the header files you might need for building your own programs from those used internally. To put all relevant files in a more convenient arrangement, run the next step. Install all relevant files created by the build:
cmake --install . --prefix c:\\opt -v\nC:\\opt as given above, this would be: C:\\opt\\include\\openblas for the header files, C:\\opt\\bin for the libopenblas.dll shared library,C:\\opt\\lib for the static library, andC:\\opt\\share holds various support files that enable other cmake-based build scripts to find OpenBLAS automatically.Change in complex types for Visual Studio 2017 and up
In newer Visual Studio versions, Microsoft has changed how it handles complex types. Even when using a precompiled version of OpenBLAS, you might need to define LAPACK_COMPLEX_CUSTOM in order to define complex types properly for MSVC. For example, some variant of the following might help:
#if defined(_MSC_VER)\n #include <complex.h>\n #define LAPACK_COMPLEX_CUSTOM\n #define lapack_complex_float _Fcomplex\n #define lapack_complex_double _Dcomplex\n#endif\nFor reference, see openblas#3661, lapack#683, and this Stack Overflow question.
Building 32-bit binaries with MSVC
This method may produce binaries which demonstrate significantly lower performance than those built with the other methods. The Visual Studio compiler does not support the dialect of assembly used in the cpu-specific optimized files, so only the \"generic\" TARGET which is written in pure C will get built. For the same reason it is not possible (and not necessary) to use -DDYNAMIC_ARCH=ON in a Visual Studio build. You may consider building for the 32-bit architecture using the GNU (MinGW) ABI instead.
To generate Visual Studio solution files, ensure CMake is installed and then run:
# Do this from Powershell so cmake can find visual studio\ncmake -G \"Visual Studio 14 Win64\" -DCMAKE_BUILD_TYPE=Release .\nTo then build OpenBLAS using those solution files from within Visual Studio, we also need Perl. Please install it and ensure it's on the PATH (see, e.g., this Stack Overflow question for how).
If you build from within Visual Studio, the dependencies may not be automatically configured: if you try to build libopenblas directly, it may fail with a message saying that some .obj files aren't found. If this happens, you can work around the problem by building the projects that libopenblas depends on before building libopenblas itself.
OpenBLAS can be built targeting Universal Windows Platform (UWP) like this:
Remove the generated CMakeCache.txt and the CMakeFiles directory from the OpenBLAS source directory, then re-run CMake with the following options:
# do this to build UWP compatible solution files\ncmake -G \"Visual Studio 14 Win64\" -DCMAKE_SYSTEM_NAME=WindowsStore -DCMAKE_SYSTEM_VERSION=\"10.0\" -DCMAKE_SYSTEM_PROCESSOR=AMD64 -DVS_WINRT_COMPONENT=TRUE -DCMAKE_BUILD_TYPE=Release .\nNote
The resulting library from building with MinGW as described below can be used in Visual Studio, but it can only be linked dynamically. This configuration has not been thoroughly tested and should be considered experimental.
To build OpenBLAS on Windows with MinGW:
gfortran package as well (unless you really want to build the BLAS part of OpenBLAS only) and check that gcc and gfortran are the same version. In addition, please install MSYS2 with MinGW.make. OpenBLAS will detect the compiler and CPU automatically. libopenblas.a and the shared library libopenblas.dll in the folder. You can type make PREFIX=/your/installation/path install to install the library to a certain location.Note that OpenBLAS will generate the import library libopenblas.dll.a for libopenblas.dll by default.
If you want to generate Windows-native PDB files from a MinGW build, you can use the cv2pdb tool to do so.
To then use the built OpenBLAS shared library in Visual Studio:
OPENBLAS_TOP_DIR/libopenblas.dll.a) and the shared library (libopenblas.dll) into the same folder (this must be the folder of your project that is going to use the BLAS library. You may need to add libopenblas.dll.a to the linker input list: properties->Linker->Input).cblas.h in /your/installation/path/include in Visual Studio. Please see openblas#95 for more details.Limitations of using the MinGW build within Visual Studio
A fully functional native OpenBLAS for WoA that can be built as both a static and dynamic library using LLVM toolchain and Visual Studio 2022. Before starting to build, make sure that you have installed Visual Studio 2022 on your ARM device, including the \"Desktop Development with C++\" component (that contains the cmake tool). (Note that you can use the free \"Visual Studio 2022 Community Edition\" for this task. In principle it would be possible to build with VisualStudio alone, but using the LLVM toolchain enables native compilation of the Fortran sources of LAPACK and of all the optimized assembly files, which VisualStudio cannot handle on its own)
Clone OpenBLAS to your local machine and checkout to latest release of OpenBLAS (unless you want to build the latest development snapshot - here we are using the 0.3.28 release as the example, of course this exact version may be outdated by the time you read this)
git clone https://github.com/OpenMathLib/OpenBLAS.git\ncd OpenBLAS\ngit checkout v0.3.28\nInstall Latest LLVM toolchain for WoA:
Download the Latest LLVM toolchain for WoA from the Release page. At the time of writing, this is version 19.1.5 - be sure to select the latest release for which you can find a precompiled package whose name ends in \"-woa64.exe\" (precompiled packages usually lag a week or two behind their corresponding source release). Make sure to enable the option \u201cAdd LLVM to the system PATH for all the users\u201d Note: Make sure that the path of LLVM toolchain is at the top of Environment Variables section to avoid conflicts between the set of compilers available in the system path
From the start menu search for \u201cARM64 Native Tools Command Prompt for Visual Studio 2022\u201d Alternatively open command prompt, run the following command to activate the environment: \"C:\\Program Files\\Microsoft Visual Studio\\2022\\Community\\VC\\Auxiliary\\Build\\vcvarsarm64.bat\"
Navigate to the OpenBLAS source code directory and start building OpenBLAS by invoking Ninja:
```cmd\n cd OpenBLAS\n mkdir build\n cd build\n\n cmake .. -G Ninja -DCMAKE_BUILD_TYPE=Release -DTARGET=ARMV8 -DBINARY=64 -DCMAKE_C_COMPILER=clang-cl -DCMAKE_C_COMPILER=arm64-pc-windows-msvc -DCMAKE_ASM_COMPILER=arm64-pc-windows-msvc -DCMAKE_Fortran_COMPILER=flang-new\n\n ninja -j16\n ```\nNote: You might want to include additional options in the cmake command here. For example, the default configuration only generates a static.lib version of the library. If you prefer a DLL, you can add -DBUILD_SHARED_LIBS=ON.
Note that it is also possible to use the same setup to build OpenBLAS with Make, if you prepare Makefiles over the CMake build for some reason:
```cmd\n$ make CC=clang-cl FC=flang-new AR=\"llvm-ar\" TARGET=ARMV8 ARCH=arm64 RANLIB=\"llvm-ranlib\" MAKE=make\n```\nMicrosoft Windows has this thing called \"import libraries\". You need it for MSVC; you don't need it for MinGW because the ld linker is smart enough - however, you may still want it for some reason, so we'll describe the process for both MSVC and MinGW.
Import libraries are compiled from a list of what symbols to use, which are contained in a .def file. A .def file should be already be present in the exports directory under the top-level OpenBLAS directory after you've run a build. In your shell, move to this directory: cd exports.
Unlike MinGW, MSVC absolutely requires an import library. Now the C ABI of MSVC and MinGW are actually identical, so linking is actually okay (any incompatibility in the C ABI would be a bug).
The import libraries of MSVC have the suffix .lib. They are generated from a .def file using MSVC's lib.exe. See the MSVC instructions.
MinGW import libraries have the suffix .a, just like static libraries. Our goal is to produce the file libopenblas.dll.a.
You need to first insert a line LIBRARY libopenblas.dll in libopenblas.def:
cat <(echo \"LIBRARY libopenblas.dll\") libopenblas.def > libopenblas.def.1\nmv libopenblas.def.1 libopenblas.def\nNow the .def file probably looks like:
LIBRARY libopenblas.dll\nEXPORTS\n caxpy=caxpy_ @1\n caxpy_=caxpy_ @2\n ...\ndlltool -d libopenblas.def -l libopenblas.dll.a Again, there is basically no point in making an import library for use in MinGW. It actually slows down linking.
"},{"location":"install/#android","title":"Android","text":"To build OpenBLAS for Android, you will need the following tools installed on your machine:
The next two sections below describe how to build with Clang for ARMV7 and ARMV8 targets, respectively. The same basic principles as described below for ARMV8 should also apply to building an x86 or x86-64 version (substitute something like NEHALEM for the target instead of ARMV8, and replace all the aarch64 in the toolchain paths with x86 or x96_64 as appropriate).
Historic note
Since NDK version 19, the default toolchain is provided as a standalone toolchain, so building one yourself following building a standalone toolchain should no longer be necessary.
"},{"location":"install/#building-for-armv7","title":"Building for ARMV7","text":"# Set path to ndk-bundle\nexport NDK_BUNDLE_DIR=/path/to/ndk-bundle\n\n# Set the PATH to contain paths to clang and arm-linux-androideabi-* utilities\nexport PATH=${NDK_BUNDLE_DIR}/toolchains/arm-linux-androideabi-4.9/prebuilt/linux-x86_64/bin:${NDK_BUNDLE_DIR}/toolchains/llvm/prebuilt/linux-x86_64/bin:$PATH\n\n# Set LDFLAGS so that the linker finds the appropriate libgcc\nexport LDFLAGS=\"-L${NDK_BUNDLE_DIR}/toolchains/arm-linux-androideabi-4.9/prebuilt/linux-x86_64/lib/gcc/arm-linux-androideabi/4.9.x\"\n\n# Set the clang cross compile flags\nexport CLANG_FLAGS=\"-target arm-linux-androideabi -marm -mfpu=vfp -mfloat-abi=softfp --sysroot ${NDK_BUNDLE_DIR}/platforms/android-23/arch-arm -gcc-toolchain ${NDK_BUNDLE_DIR}/toolchains/arm-linux-androideabi-4.9/prebuilt/linux-x86_64/\"\n\n#OpenBLAS Compile\nmake TARGET=ARMV7 ONLY_CBLAS=1 AR=ar CC=\"clang ${CLANG_FLAGS}\" HOSTCC=gcc ARM_SOFTFP_ABI=1 -j4\nOn macOS, it may also be necessary to give the complete path to the ar utility in the make command above, like so:
AR=${NDK_BUNDLE_DIR}/toolchains/arm-linux-androideabi-4.9/prebuilt/darwin-x86_64/bin/arm-linux-androideabi-gcc-ar\nmalformed archive header name at 8 when the native macOS ar command was invoked instead. Note that with recent NDK versions, the AR tool may be named llvm-ar rather than what is assumed above."},{"location":"install/#building-for-armv8","title":"Building for ARMV8","text":"# Set path to ndk-bundle\nexport NDK_BUNDLE_DIR=/path/to/ndk-bundle/\n\n# Export PATH to contain directories of clang and aarch64-linux-android-* utilities\nexport PATH=${NDK_BUNDLE_DIR}/toolchains/aarch64-linux-android-4.9/prebuilt/linux-x86_64/bin/:${NDK_BUNDLE_DIR}/toolchains/llvm/prebuilt/linux-x86_64/bin:$PATH\n\n# Setup LDFLAGS so that loader can find libgcc and pass -lm for sqrt\nexport LDFLAGS=\"-L${NDK_BUNDLE_DIR}/toolchains/aarch64-linux-android-4.9/prebuilt/linux-x86_64/lib/gcc/aarch64-linux-android/4.9.x -lm\"\n\n# Setup the clang cross compile options\nexport CLANG_FLAGS=\"-target aarch64-linux-android --sysroot ${NDK_BUNDLE_DIR}/platforms/android-23/arch-arm64 -gcc-toolchain ${NDK_BUNDLE_DIR}/toolchains/aarch64-linux-android-4.9/prebuilt/linux-x86_64/\"\n\n# Compile\nmake TARGET=ARMV8 ONLY_CBLAS=1 AR=ar CC=\"clang ${CLANG_FLAGS}\" HOSTCC=gcc -j4\nTARGET=CORTEXA57 in place of ARMV8 will pick up better optimized routines. Implementations for the CORTEXA57 target are compatible with all other ARMV8 targets. Note: for NDK 23b, something as simple as:
export PATH=/opt/android-ndk-r23b/toolchains/llvm/prebuilt/linux-x86_64/bin/:$PATH\nmake HOSTCC=gcc CC=/opt/android-ndk-r23b/toolchains/llvm/prebuilt/linux-x86_64/bin/aarch64-linux-android31-clang ONLY_CBLAS=1 TARGET=ARMV8\nllvm-ar) is also necessary. Alternative build script for 3 architectures This script will build OpenBLAS for 3 architecture (ARMV7, ARMV8, X86) and install them to /opt/OpenBLAS/lib. Of course you can also copy only the section that is of interest to you - also notice that the AR= line may need adapting to the name of the ar tool provided in your $TOOLCHAIN/bin - for example llvm-ar in some recent NDK versions. It was tested on macOS with NDK version 21.3.6528147.
export NDK=YOUR_PATH_TO_SDK/Android/sdk/ndk/21.3.6528147\nexport TOOLCHAIN=$NDK/toolchains/llvm/prebuilt/darwin-x86_64\n\nmake clean\nmake \\\n TARGET=ARMV7 \\\n ONLY_CBLAS=1 \\\n CC=\"$TOOLCHAIN\"/bin/armv7a-linux-androideabi21-clang \\\n AR=\"$TOOLCHAIN\"/bin/arm-linux-androideabi-ar \\\n HOSTCC=gcc \\\n ARM_SOFTFP_ABI=1 \\\n -j4\nsudo make install\n\nmake clean\nmake \\\n TARGET=CORTEXA57 \\\n ONLY_CBLAS=1 \\\n CC=$TOOLCHAIN/bin/aarch64-linux-android21-clang \\\n AR=$TOOLCHAIN/bin/aarch64-linux-android-ar \\\n HOSTCC=gcc \\\n-j4\nsudo make install\n\nmake clean\nmake \\\n TARGET=ATOM \\\n ONLY_CBLAS=1 \\\nCC=\"$TOOLCHAIN\"/bin/i686-linux-android21-clang \\\nAR=\"$TOOLCHAIN\"/bin/i686-linux-android-ar \\\nHOSTCC=gcc \\\nARM_SOFTFP_ABI=1 \\\n-j4\nsudo make install\n\n## This will build for x86_64 \nmake clean\nmake \\\n TARGET=ATOM BINARY=64\\\nONLY_CBLAS=1 \\\nCC=\"$TOOLCHAIN\"/bin/x86_64-linux-android21-clang \\\nAR=\"$TOOLCHAIN\"/bin/x86_64-linux-android-ar \\\nHOSTCC=gcc \\\nARM_SOFTFP_ABI=1 \\\n-j4\nsudo make install\nAs none of the current developers uses iOS, the following instructions are what was found to work in our Azure CI setup, but as far as we know this builds a fully working OpenBLAS for this platform.
Go to the directory where you unpacked OpenBLAS,and enter the following commands:
CC=/Applications/Xcode_12.4.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin/clang\n\nCFLAGS= -O2 -Wno-macro-redefined -isysroot /Applications/Xcode_12.4.app/Contents/Developer/Platforms/iPhoneOS.platform/Developer/SDKs/iPhoneOS14.4.sdk -arch arm64 -miphoneos-version-min=10.0\n\nmake TARGET=ARMV8 DYNAMIC_ARCH=1 NUM_THREADS=32 HOSTCC=clang NOFORTRAN=1\nMIN_IOS_VERSION as necessary for your installation. E.g., change the version number to the minimum iOS version you want to target and execute this file to build the library."},{"location":"install/#harmonyos","title":"HarmonyOS","text":"For this target you will need the cross-compiler toolchain package by Huawei, which contains solutions for both Windows and Linux. Only the Linux-based toolchain has been tested so far, but the following instructions may apply similarly to Windows:
Download https://repo.huaweicloud.com/harmonyos/os/4.1.1-Release/ohos-sdk-windows_linux-public.tar.gz (or whatever newer version may be available in the future). Use tar xvf ohos-sdk-windows_linux_public.tar.gz to unpack it somewhere on your system. This will create a folder named \"ohos-sdk\" with subfolders \"linux\" and \"windows\". In the linux one you will find a ZIP archive named \"native-linux-x64-4.1.7.8-Release.zip\" - you need to unzip this where you want to install the cross-compiler, for example in /opt/ohos-sdk.
In the directory where you unpacked OpenBLAS, create a build directory for cmake, and change into it :
mkdir build\ncd build\ncmake that came with the SDK, and specify the location of its toolchain file as a cmake option. Also set the build target for OpenBLAS to ARMV8 and specify NOFORTRAN=1 (at least as of version 4.1.1, the SDK contains no Fortran compiler): /opt/ohos-sdk/linux/native/build-tools/cmake/bin/cmake -DCMAKE_TOOLCHAIN_FILE=/opt/ohos-sdk/linux/native/build/cmake/ohos.toolchain.cmake \\\n -DOHOS_ARCH=\"arm64-v8a\" -DTARGET=ARMV8 -DNOFORTRAN=1 ..\n/opt/ohos-sdk/linux/native/build-tools/cmake/bin/cmake --build .\nFor MIPS targets you will need latest toolchains:
You can use following commandlines for builds:
IMG_TOOLCHAIN_DIR={full IMG GNU/Linux Toolchain path including \"bin\" directory -- for example, /opt/linux_toolchain/bin}\nIMG_GCC_PREFIX=mips-img-linux-gnu\nIMG_TOOLCHAIN=${IMG_TOOLCHAIN_DIR}/${IMG_GCC_PREFIX}\n\n# I6400 Build (n32):\nmake BINARY=32 BINARY32=1 CC=$IMG_TOOLCHAIN-gcc AR=$IMG_TOOLCHAIN-ar FC=\"$IMG_TOOLCHAIN-gfortran -EL -mabi=n32\" RANLIB=$IMG_TOOLCHAIN-ranlib HOSTCC=gcc CFLAGS=\"-EL\" FFLAGS=$CFLAGS LDFLAGS=$CFLAGS TARGET=I6400\n\n# I6400 Build (n64):\nmake BINARY=64 BINARY64=1 CC=$IMG_TOOLCHAIN-gcc AR=$IMG_TOOLCHAIN-ar FC=\"$IMG_TOOLCHAIN-gfortran -EL\" RANLIB=$IMG_TOOLCHAIN-ranlib HOSTCC=gcc CFLAGS=\"-EL\" FFLAGS=$CFLAGS LDFLAGS=$CFLAGS TARGET=I6400\n\n# P6600 Build (n32):\nmake BINARY=32 BINARY32=1 CC=$IMG_TOOLCHAIN-gcc AR=$IMG_TOOLCHAIN-ar FC=\"$IMG_TOOLCHAIN-gfortran -EL -mabi=n32\" RANLIB=$IMG_TOOLCHAIN-ranlib HOSTCC=gcc CFLAGS=\"-EL\" FFLAGS=$CFLAGS LDFLAGS=$CFLAGS TARGET=P6600\n\n# P6600 Build (n64):\nmake BINARY=64 BINARY64=1 CC=$IMG_TOOLCHAIN-gcc AR=$IMG_TOOLCHAIN-ar FC=\"$IMG_TOOLCHAIN-gfortran -EL\" RANLIB=$IMG_TOOLCHAIN-ranlib HOSTCC=gcc CFLAGS=\"-EL\" FFLAGS=\"$CFLAGS\" LDFLAGS=\"$CFLAGS\" TARGET=P6600\n\nMTI_TOOLCHAIN_DIR={full MTI GNU/Linux Toolchain path including \"bin\" directory -- for example, /opt/linux_toolchain/bin}\nMTI_GCC_PREFIX=mips-mti-linux-gnu\nMTI_TOOLCHAIN=${IMG_TOOLCHAIN_DIR}/${IMG_GCC_PREFIX}\n\n# P5600 Build:\n\nmake BINARY=32 BINARY32=1 CC=$MTI_TOOLCHAIN-gcc AR=$MTI_TOOLCHAIN-ar FC=\"$MTI_TOOLCHAIN-gfortran -EL\" RANLIB=$MTI_TOOLCHAIN-ranlib HOSTCC=gcc CFLAGS=\"-EL\" FFLAGS=$CFLAGS LDFLAGS=$CFLAGS TARGET=P5600\nYou will need to install the following tools from the FreeBSD ports tree:
To compile run the command:
$ gmake CC=gcc FC=gfortran\nCortex-M is a widely used microcontroller that is present in a variety of industrial and consumer electronics. A common variant of the Cortex-M is the STM32F4xx series. Here, we will give instructions for building for that series.
First, install the embedded Arm GCC compiler from the Arm website. Then, create the following toolchain.cmake file:
set(CMAKE_SYSTEM_NAME Generic)\nset(CMAKE_SYSTEM_PROCESSOR arm)\n\nset(CMAKE_C_COMPILER \"arm-none-eabi-gcc.exe\")\nset(CMAKE_CXX_COMPILER \"arm-none-eabi-g++.exe\")\n\nset(CMAKE_EXE_LINKER_FLAGS \"--specs=nosys.specs\" CACHE INTERNAL \"\")\n\nset(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM NEVER)\nset(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY ONLY)\nset(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)\nset(CMAKE_FIND_ROOT_PATH_MODE_PACKAGE ONLY)\nThen build OpenBLAS with:
$ cmake .. -G Ninja -DCMAKE_C_COMPILER=arm-none-eabi-gcc -DCMAKE_TOOLCHAIN_FILE:PATH=\"toolchain.cmake\" -DNOFORTRAN=1 -DTARGET=ARMV5 -DEMBEDDED=1\nIn your embedded application, the following functions need to be provided for OpenBLAS to work correctly:
void free(void* ptr);\nvoid* malloc(size_t size);\nNote
If you are developing for an embedded platform, it is your responsibility to make sure that the device has sufficient memory for malloc calls. Libmemory provides one implementation of malloc for embedded platforms.
If the OpenBLAS DLLs are not linked correctly, you may see an error like \"The application was unable to start correctly (0xc000007b)\", which typically indicates a mismatch between 32-bit and 64-bit libraries.\u00a0\u21a9
This user manual covers compiling OpenBLAS itself, linking your code to OpenBLAS, example code to use the C (CBLAS) and Fortran (BLAS) APIs, and some troubleshooting tips. Compiling OpenBLAS is optional, since you may be able to install with a package manager.
Note
The OpenBLAS documentation does not contain API reference documentation for BLAS or LAPACK, since these are standardized APIs, the documentation for which can be found in other places. If you want to understand every BLAS and LAPACK function and definition, we recommend reading the Netlib BLAS and Netlib LAPACK documentation.
OpenBLAS does contain a limited number of functions that are non-standard, these are documented at OpenBLAS extension functions.
"},{"location":"user_manual/#compiling-openblas","title":"Compiling OpenBLAS","text":""},{"location":"user_manual/#normal-compile","title":"Normal compile","text":"The default way to build and install OpenBLAS from source is with Make:
make # add `-j4` to compile in parallel with 4 processes\nmake install\nBy default, the CPU architecture is detected automatically when invoking make, and the build is optimized for the detected CPU. To override the autodetection, use the TARGET flag:
# `make TARGET=xxx` sets target CPU: e.g. for an Intel Nehalem CPU:\nmake TARGET=NEHALEM\nTargetList.txt in the root of the repository."},{"location":"user_manual/#cross-compile","title":"Cross compile","text":"For a basic cross-compilation with Make, three steps need to be taken:
CC and FC environment variables to select the cross toolchains for C and Fortran.HOSTCC environment variable to select the host C compiler (i.e. the regular C compiler for the machine on which you are invoking the build).TARGET explicitly to the CPU architecture on which the produced OpenBLAS binaries will be used.Compile the library for ARM Cortex-A9 linux on an x86-64 machine (note: install only gnueabihf versions of the cross toolchain - see this issue comment for why):
make CC=arm-linux-gnueabihf-gcc FC=arm-linux-gnueabihf-gfortran HOSTCC=gcc TARGET=CORTEXA9\nCompile OpenBLAS for a loongson3a CPU on an x86-64 machine:
make BINARY=64 CC=mips64el-unknown-linux-gnu-gcc FC=mips64el-unknown-linux-gnu-gfortran HOSTCC=gcc TARGET=LOONGSON3A\nCompile OpenBLAS for loongson3a CPU with the loongcc (based on Open64) compiler on an x86-64 machine:
make CC=loongcc FC=loongf95 HOSTCC=gcc TARGET=LOONGSON3A CROSS=1 CROSS_SUFFIX=mips64el-st-linux-gnu- NO_LAPACKE=1 NO_SHARED=1 BINARY=32\nAdd DEBUG=1 to your build command, e.g.:
make DEBUG=1\nNote
Installing to a directory is optional; it is also possible to use the shared or static libraries directly from the build directory.
Use make install with the PREFIX flag to install to a specific directory:
make install PREFIX=/path/to/installation/directory\nThe default directory is /opt/OpenBLAS.
Important
Note that any flags passed to make during build should also be passed to make install to circumvent any install errors, i.e. some headers not being copied over correctly.
For more detailed information on building/installing from source, please read the Installation Guide.
"},{"location":"user_manual/#linking-to-openblas","title":"Linking to OpenBLAS","text":"OpenBLAS can be used as a shared or a static library.
"},{"location":"user_manual/#link-a-shared-library","title":"Link a shared library","text":"The shared library is normally called libopenblas.so, but not that the name may be different as a result of build flags used or naming choices by a distro packager (see [distributing.md] for details). To link a shared library named libopenblas.so, the flag -lopenblas is needed. To find the OpenBLAS headers, a -I/path/to/includedir is needed. And unless the library is installed in a directory that the linker searches by default, also -L and -Wl,-rpath flags are needed. For a source file test.c (e.g., the example code under Call CBLAS interface further down), the shared library can then be linked with:
gcc -o test test.c -I/your_path/OpenBLAS/include/ -L/your_path/OpenBLAS/lib -Wl,-rpath,/your_path/OpenBLAS/lib -lopenblas\nThe -Wl,-rpath,/your_path/OpenBLAS/lib linker flag can be omitted if you ran ldconfig to update linker cache, put /your_path/OpenBLAS/lib in /etc/ld.so.conf or a file in /etc/ld.so.conf.d, or installed OpenBLAS in a location that is part of the ld.so default search path (usually /lib, /usr/lib and /usr/local/lib). Alternatively, you can set the environment variable LD_LIBRARY_PATH to point to the folder that contains libopenblas.so. Otherwise, the build may succeed but at runtime loading the library will fail with a message like:
cannot open shared object file: no such file or directory\nMore flags may be needed, depending on how OpenBLAS was built:
libopenblas is multi-threaded, please add -lpthread.-lgfortran (other Fortran libraries may also be needed, e.g. -lquadmath). Note that if you only make calls to LAPACKE routines, i.e. your code has #include \"lapacke.h\" and makes calls to methods like LAPACKE_dgeqrf, then -lgfortran is not needed.Tip
Usually a pkg-config file (e.g., openblas.pc) is installed together with a libopenblas shared library. pkg-config is a tool that will tell you the exact flags needed for linking. For example:
$ pkg-config --cflags openblas\n-I/usr/local/include\n$ pkg-config --libs openblas\n-L/usr/local/lib -lopenblas\nLinking a static library is simpler - add the path to the static OpenBLAS library to the compile command:
gcc -o test test.c /your/path/libopenblas.a\nThis example shows calling cblas_dgemm in C:
#include <cblas.h>\n#include <stdio.h>\n\nvoid main()\n{\n int i=0;\n double A[6] = {1.0,2.0,1.0,-3.0,4.0,-1.0}; \n double B[6] = {1.0,2.0,1.0,-3.0,4.0,-1.0}; \n double C[9] = {.5,.5,.5,.5,.5,.5,.5,.5,.5}; \n cblas_dgemm(CblasColMajor, CblasNoTrans, CblasTrans,3,3,2,1,A, 3, B, 3,2,C,3);\n\n for(i=0; i<9; i++)\n printf(\"%lf \", C[i]);\n printf(\"\\n\");\n}\nTo compile this file, save it as test_cblas_dgemm.c and then run:
gcc -o test_cblas_open test_cblas_dgemm.c -I/your_path/OpenBLAS/include/ -L/your_path/OpenBLAS/lib -lopenblas -lpthread -lgfortran\ntest_cblas_open executable."},{"location":"user_manual/#call-blas-fortran-interface","title":"Call BLAS Fortran interface","text":"This example shows calling the dgemm Fortran interface in C:
#include \"stdio.h\"\n#include \"stdlib.h\"\n#include \"sys/time.h\"\n#include \"time.h\"\n\nextern void dgemm_(char*, char*, int*, int*,int*, double*, double*, int*, double*, int*, double*, double*, int*);\n\nint main(int argc, char* argv[])\n{\n int i;\n printf(\"test!\\n\");\n if(argc<4){\n printf(\"Input Error\\n\");\n return 1;\n }\n\n int m = atoi(argv[1]);\n int n = atoi(argv[2]);\n int k = atoi(argv[3]);\n int sizeofa = m * k;\n int sizeofb = k * n;\n int sizeofc = m * n;\n char ta = 'N';\n char tb = 'N';\n double alpha = 1.2;\n double beta = 0.001;\n\n struct timeval start,finish;\n double duration;\n\n double* A = (double*)malloc(sizeof(double) * sizeofa);\n double* B = (double*)malloc(sizeof(double) * sizeofb);\n double* C = (double*)malloc(sizeof(double) * sizeofc);\n\n srand((unsigned)time(NULL));\n\n for (i=0; i<sizeofa; i++)\n A[i] = i%3+1;//(rand()%100)/10.0;\n\n for (i=0; i<sizeofb; i++)\n B[i] = i%3+1;//(rand()%100)/10.0;\n\n for (i=0; i<sizeofc; i++)\n C[i] = i%3+1;//(rand()%100)/10.0;\n //#if 0\n printf(\"m=%d,n=%d,k=%d,alpha=%lf,beta=%lf,sizeofc=%d\\n\",m,n,k,alpha,beta,sizeofc);\n gettimeofday(&start, NULL);\n dgemm_(&ta, &tb, &m, &n, &k, &alpha, A, &m, B, &k, &beta, C, &m);\n gettimeofday(&finish, NULL);\n\n duration = ((double)(finish.tv_sec-start.tv_sec)*1000000 + (double)(finish.tv_usec-start.tv_usec)) / 1000000;\n double gflops = 2.0 * m *n*k;\n gflops = gflops/duration*1.0e-6;\n\n FILE *fp;\n fp = fopen(\"timeDGEMM.txt\", \"a\");\n fprintf(fp, \"%dx%dx%d\\t%lf s\\t%lf MFLOPS\\n\", m, n, k, duration, gflops);\n fclose(fp);\n\n free(A);\n free(B);\n free(C);\n return 0;\n}\nTo compile this file, save it as time_dgemm.c and then run:
gcc -o time_dgemm time_dgemm.c /your/path/libopenblas.a -lpthread\n./time_dgemm <m> <n> <k>, with m, n, and k input parameters to the time_dgemm executable. Note
When calling the Fortran interface from C, you have to deal with symbol name differences caused by compiler conventions. That is why the dgemm_ function call in the example above has a trailing underscore. This is what it looks like when using gcc/gfortran, however such details may change for different compilers. Hence it requires extra support code. The CBLAS interface may be more portable when writing C code.
When writing code that needs to be portable and work across different platforms and compilers, the above code example is not recommended for usage. Instead, we advise looking at how OpenBLAS (or BLAS in general, since this problem isn't specific to OpenBLAS) functions are called in widely used projects like Julia, SciPy, or R.
"},{"location":"user_manual/#troubleshooting","title":"Troubleshooting","text":"BIGNUMA=1.NO_AFFINITY=1 in Makefile.rule.make test is known to fail with a pthread_create error and an EAGAIN error code. However, it will be OK when you run the same testcase in a shell.