perf-libs-sparse

A library of high-performance sparse linear algebra functions targeting AArch64.

perf-libs-sparse provides sparse linear algebra kernels and interfaces for AArch64 systems, with C and Fortran support, configurable LP64/ILP64 interfaces, and optional OpenMP parallelism. The project may be built on Linux, Windows or macOS using CMake.

Table of Contents

• Quick Start
• Installation
• Usage
• Contributing
• License

Quick Start

git clone https://github.com/arm/perf-libs-sparse.git
cd perf-libs-sparse
python3 -m pip install --user mako
CC=gcc CXX=g++ FC=gfortran cmake -S . -B build
cmake --build build --parallel
ctest --test-dir build --output-on-failure

Installation

Prerequisites

• CMake 3.20 or newer
• A C compiler and C++ compiler with C++17 support
• A Fortran compiler with Fortran 2003 support
• Note: GNU and LLVM are the recommended compiler suites
• Python 3 with the mako package available to CMake
• Arm Performance Libraries with $ARMPL_DIR set, or a system OpenBLAS installation if SPARSE_USE_ARMPL=OFF (perf-libs-sparse has a dependency on some CBLAS functions)

Configure, Build, and Install

python3 -m pip install --user mako
CC=clang CXX=clang++ FC=flang cmake -S . -B build
cmake --build build --parallel
cmake --install build --prefix ./_install

Useful build options:

• -DSPARSE_BUILD_SHARED=OFF to disable the shared library build
• -DSPARSE_BUILD_STATIC=OFF to disable the static library build
• -DSPARSE_ENABLE_FORTRAN=OFF to disable the Fortran interface and module build
• -DSPARSE_USE_ARMPL=OFF to use a system OpenBLAS backend instead of ArmPL
• -DSPARSE_ENABLE_OPENMP=ON to build OpenMP variants
• -DSPARSE_ENABLE_ASAN=ON to enable AddressSanitizer
• -DSPARSE_ENABLE_COVERAGE=ON to add compiler coverage instrumentation for supported GNU/Clang toolchains
• -DSPARSE_ENABLE_INT64=ON to build with ILP64 interfaces
• -DSPARSE_ENABLE_ARM64EC=ON to target Arm64EC on Windows
• -DSPARSE_ENABLE_TESTS=OFF to disable building the test executables
• -DSPARSE_ENABLE_SVE=ON to build with SVE kernels

Usage

Run the full test suite:

ctest --test-dir build --output-on-failure

Run an individual test executable:

./build/bin/spmv_basic_d_test

Consume the installed package from another CMake project:

find_package(perflibs_sparse_lp64 REQUIRED CONFIG)
target_link_libraries(my_target PRIVATE perflibs_sparse_lp64::sparse_shared)

This is the default package for most consumers.

If you specifically need the ILP64 variant, use perflibs_sparse_ilp64 instead. For find_package(... CONFIG), CMake can also use the standard <PackageName>_DIR hint variable, so perflibs_sparse_lp64_DIR or perflibs_sparse_ilp64_DIR may be set to the directory containing the corresponding *Config.cmake file.

Contributing

Contributions are welcome. Please follow the project guidelines in CONTRIBUTING.md.

License

This project is licensed under the terms described in LICENSE.