使用 Homebrew, apt, dnf, MacPorts, Nix, pacman 安装 clblast

项目链接

• 主页
• 上游文档
• 软件包管理器页面

命令和别名

• clblast_tuner_copy_fast
• clblast_tuner_copy_pad
• clblast_tuner_invert
• clblast_tuner_routine_xgemm
• clblast_tuner_routine_xtrsv
• clblast_tuner_transpose_fast
• clblast_tuner_transpose_pad
• clblast_tuner_xaxpy
• clblast_tuner_xconvgemm
• clblast_tuner_xdot
• clblast_tuner_xgemm
• clblast_tuner_xgemm_direct
• clblast_tuner_xgemv
• clblast_tuner_xger

项目历史

The changelog starts with version 0.1.0 as the initial preview release on GitHub, supporting a small set of level-1 and level-3 routines. Later pre-1.0 releases added broader BLAS coverage, a C API, Travis CI, MSVC support, thread safety, testing against clBLAS and CPU BLAS libraries, and tuned device parameters.

CLBlast reached 1.0.0 after adding correctness fixes, Mesa Clover and ROCm support, external performance reports, and more tuned parameters. The README cites a 2017 GTC presentation, an IWOCL 2018 presentation, and the paper 'CLBlast: A Tuned OpenCL BLAS Library' as deeper background.

采用历史

The upstream changelog says version 1.4.0 added Ubuntu PPA and macOS Homebrew package-manager availability, and the batch metadata lists Homebrew, Debian, Fedora, MacPorts, Nix, Arch, and Ubuntu package names. Adoption was driven by users needing a tunable OpenCL BLAS on devices not well served by fixed vendor libraries.

The README asks users with poor out-of-the-box performance to run tuners and contribute results, which made the project partly a packaged library and partly a community database of device-specific OpenCL tuning.

使用方式

CLBlast users compile the library with CMake, include either clblast.h or clblast_c.h, pass OpenCL buffers to routines, and optionally build samples, benchmarks, tests, and tuners. The API is intentionally close to Netlib BLAS and cuBLAS/clBLAS conventions while avoiding a library initialization step.

为什么软件包爱好者会关心

CLBlast is package-nerd bait because it combines portable C++ packaging, OpenCL ICD dependencies, optional CUDA backend support, Python bindings, tuners, samples, benchmarks, and a built-in tuning database. It also offers a clean contrast with clBLAS: same problem domain, but newer, more tunable, and more explicit about cross-vendor device coverage.

时间线

• 2017: CLBlast is presented at NVIDIA GTC and the first arXiv version of the CLBlast paper appears.
• 2018: CLBlast is presented at IWOCL and the paper is published as 'CLBlast: A Tuned OpenCL BLAS Library'.
• 1.0.0: The project reaches a stable major release with ROCm/Mesa support and external performance reports.
• 1.4.0: Ubuntu PPA and macOS Homebrew packaging are added upstream.
• 1.7.0: Parallel kernel compilation for faster tuning and CMake 4.0 compatibility are added.

Related projects

• clBLAS is the older OpenCL BLAS library that CLBlast explicitly compares against and offers migration-compatible APIs for.
• cuBLAS is the NVIDIA CUDA BLAS library used as a contrast in the README.
• CLTune inspired the included auto-tuner according to the README's paper references.
• Netlib BLAS provides the routine model, while OpenCL provides the device API.

来源

• https://github.com/CNugteren/CLBlast#readme
• https://raw.githubusercontent.com/CNugteren/CLBlast/master/CHANGELOG
• https://cnugteren.github.io/clblast/clblast.html
• https://arxiv.org/abs/1705.05249