百科.dev
全部条目在线工具AI 编程趋势榜开源项目技术资讯提交条目
登录
< 返回工具列表
B

BabelStream

> 编程语言
开源

STREAM, 适用于使用多种编程模型编写的众多设备

372 stars0 点赞0 次浏览
访问官网GitHub

工具介绍

STREAM, 适用于使用多种编程模型编写的众多设备

BabelStream

Measure memory transfer rates to/from global device memory on GPUs. This benchmark is similar in spirit, and based on, the STREAM benchmark [1] for CPUs.

Unlike other GPU memory bandwidth benchmarks this does not include the PCIe transfer time.

There are multiple implementations of this benchmark in a variety of programming models.

This code was previously called GPU-STREAM.

Table of Contents

  • Programming Models
  • How is this different to STREAM?
  • Building
    • CMake
    • Spack
    • GNU Make (removed)
  • Results
  • Contributing
  • Citing
    • Other BabelStream publications

Programming Models

BabelStream is currently implemented in the following parallel programming models, listed in no particular order:

  • OpenCL
  • CUDA
  • HIP
  • OpenACC
  • OpenMP 3 and 4.5
  • C++ Parallel STL
  • Kokkos
  • RAJA
  • SYCL and SYCL2020 (USM and accessors)
  • TBB
  • Thrust (via CUDA or HIP)
  • Futhark

This project also contains implementations in alternative languages with different build systems:

  • Julia - JuliaStream.jl
  • Java - java-stream
  • Scala - scala-stream
  • Rust - rust-stream

How is this different to STREAM?

BabelStream implements the four main kernels of the STREAM benchmark (along with a dot product), but by utilising different programming models expands the platforms which the code can run beyond CPUs.

The key differences from STREAM are that:

  • the arrays are allocated on the heap
  • the problem size is unknown at compile time
  • wider platform and programming model support

With stack arrays of known size at compile time, the compiler is able to align data and issue optimal instructions (such as non-temporal stores, remove peel/remainder vectorisation loops, etc.). But this information is not typically available in real HPC codes today, where the problem size is read from the user at runtime.

BabelStream therefore provides a measure of what memory bandwidth performance can be attained (by a particular programming model) if you follow today's best parallel programming best practice.

BabelStream also includes the nstream kernel from the Parallel Research Kernels (PRK) project, available on GitHub. Details about PRK can be found in the following references:

  • Van der Wijngaart, Rob F., and Timothy G. Mattson. The parallel research kernels. IEEE High Performance Extreme Computing Conference (HPEC). IEEE, 2014.

  • R. F. Van der Wijngaart, A. Kayi, J. R. Hammond, G. Jost, T. St. John, S. Sridharan, T. G. Mattson, J. Abercrombie, and J. Nelson. Comparing runtime systems with exascale ambitions using the Parallel Research Kernels. ISC 2016, DOI: 10.1007/978-3-319-41321-1_17.

  • Jeff R. Hammond and Timothy G. Mattson. Evaluating data parallelism in C++ using the Parallel Research Kernels. IWOCL 2019, DOI: 10.1145/3318170.3318192.

Building

Drivers, compiler and software applicable to whichever implementation you would like to build against is required.

CMake

The project supports building with CMake >= 3.13.0, which can be installed without root via the official script.

Each BabelStream implementation (programming model) is built as follows:

bash
$ cd babelstream

# configure the build, build type defaults to Release
# The -DMODEL flag is required
$ cmake -Bbuild -H. -DMODEL= 

# compile
$ cmake --build build

# run executables in ./build
$ ./build/-stream

The MODEL option selects one implementation of BabelStream to build. The source for each model's implementations are located in ./src/.

Currently available models are:

omp;ocl;std-data;std-indices;std-ranges;hip;cuda;kokkos;sycl;sycl2020-acc;sycl2020-usm;acc;raja;tbb;thrust;futhark

Overriding default flags

By default, we have defined a set of optimal flags for known HPC compilers. There are assigned those to RELEASE_FLAGS, and you can override them if required.

To find out what flag each model supports or requires, simply configure while only specifying the model. For example:

…

Alternatively, refer to the CI script, which test-compiles most of the models, and see which flags are used there.

It is recommended that you delete the build directory when you change any of the build flags.

Spack

The project supports building with Spack >= 0.19.0, which can be installed without root via the official GitHub repo. The BabelStream Spack Package source code could be accessed from the link here. Each BabelStream implementation (programming model) is built as follows:

…

More detailed examples are provided in Spack README file. The MODEL variant selects one implementation of BabelStream to build.

Currently available models are:

omp;ocl;std-data;std-indices;std-ranges;hip;cuda;kokkos;sycl;sycl2020-acc;sycl2020-usm;acc;raja;tbb;thrust

GNU Make

Support for Make has been removed from 4.0 onwards. However, as the build process only involves a few source files, the required compile commands can be extracted from the CI output.

Results

Sample results can be found in the results subdirectory. Newer results are found in our Performance Portability repository.

Contributing

As of v4.0, the main branch of this repository will hold the latest released version.

The develop branch will contain unreleased features due for the next (major and/or minor) release of BabelStream. Pull Requests should be made against the develop branch.

Citing

Please cite BabelStream via this reference:

Deakin T, Price J, Martineau M, McIntosh-Smith S. Evaluating attainable memory bandwidth of parallel programming models via BabelStream. International Journal of Computational Science and Engineering. Special issue. Vol. 17, No. 3, pp. 247–262. 2018. DOI: 10.1504/IJCSE.2018.095847

Other BabelStream publications

  • Deakin T, Price J, Martineau M, McIntosh-Smith S. GPU-STREAM v2.0: Benchmarking the achievable memory bandwidth of many-core processors across diverse parallel programming models. 2016. Paper presented at P^3MA Workshop at ISC High Performance, Frankfurt, Germany. DOI: 10.1007/978- 3-319-46079-6_34

  • Deakin T, McIntosh-Smith S. GPU-STREAM: Benchmarking the achievable memory bandwidth of Graphics Processing Units. 2015. Poster session presented at IEEE/ACM SuperComputing, Austin, United States. You can view the Poster and Extended Abstract.

  • Deakin T, Price J, Martineau M, McIntosh-Smith S. GPU-STREAM: Now in 2D!. 2016. Poster session presented at IEEE/ACM SuperComputing, Salt Lake City, United States. You can view the Poster and Extended Abstract.

  • Raman K, Deakin T, Price J, McIntosh-Smith S. Improving achieved memory bandwidth from C++ codes on Intel Xeon Phi Processor (Knights Landing). IXPUG Spring Meeting, Cambridge, UK, 2017.

  • Deakin T, Price J, McIntosh-Smith S. Portable methods for measuring cache hierarchy performance. 2017. Poster sessions presented at IEEE/ACM SuperComputing, Denver, United States. You can view the Poster and Extended Abstract

[1]: McCalpin, John D., 1995: "Memory Bandwidth and Machine Balance in Current High Performance Computers", IEEE Computer Society Technical Committee on Computer Architecture (TCCA) Newsletter, December 1995.

Issues· 0 开放

查看全部 Issues在 GitHub 打开

暂无开放 Issues,或尚未同步最近议题。

> 标签

C++benchmarkcudagpgpugpu

暂无评论,来聊聊你的看法吧

> 工具信息

发布日期2026年8月1日
最后更新2026年9月17日
分类编程语言
定价开源

> 相关工具

T
TypeScript
JavaScript 的超集,为前端与全栈提供静态类型
P
Python
通用编程语言,广泛用于 Web、数据与 AI
G
Go
Google 推出的简洁高效系统语言