Skip to content

Latest commit

 

History

History
84 lines (61 loc) · 3.62 KB

README.md

File metadata and controls

84 lines (61 loc) · 3.62 KB
Raw

Embench™: Open Benchmarks for Embedded Platforms

This repository contains the Embench™ free and open source benchmark suite. These benchmarks are designed to test the performance of deeply embedded systems. As such they assume the presence of no OS, minimal C library support and in particular no output stream.

The rationale behind this benchmark is described in "Embench™: An Evolving Benchmark Suite for Embedded IoT Computers from an Academic-Industrial Cooperative: Towards the Long Overdue and Deserved Demise of Dhrystone" by David Patterson, Jeremy Bennett, Palmer Dabbelt, Cesare Garlati, G. S. Madhusudan and Trevor Mudge (see https://tmt.knect365.com/risc-v-workshop-zurich/agenda/2#software_embench-tm-a-free-benchmark-suite-for-embedded-computing-from-an-academic-industry-cooperative-towards-the-long-overdue-and-deserved-demise-of-dhrystone).

The benchmarks are largely derived from the Bristol/Embecosm Embedded Benchmark Suite (BEEBS, see https://beebs.mageec.org/), which in turn draws its material from various earlier projects. A full description and user manual is in the doc directory.

Stable benchmark versions

The following git tags may be used to select the version of the repository for a stable release.

  • embench-0.5
  • embench-1.0

Using the benchmarks

The benchmarks can be used to yield a single consistent score for the performance of a platform and its compiler tool chain. The mechanism for this is described in the user manual.

  • The benchmarks should all compile to fit in 64kB of program space and use no more than 64kB of RAM

    • Note. An earlier version tried to limit RAM to 16kB, but this proved too restrictive.

  • The measurement of execution performance is designed to use "hot" caches. Thus each benchmark executes its entire code several times, before starting a timing run.

  • Execution runs are scaled to take approximately 4 second of CPU time. This is large enough to be accurately measured, yet means all 19 benchmarks, including cache warm up can be run in a few minutes.

  • To facilitate execution on machines of different performance, the tests are scaled by the clock speed of the processor.

  • The benchmarks are designed to be run on either real or simulated hardware. However for meaningful execution performance results any simulated hardware must be strictly cycle accurate.

Structure of the repository

The top level directory contains Python scripts to build and execute the benchmarks. The following are the key top level directories.

  • config: containing a directory for each architecture supported, and within that directory subdirectories for board and cpu descriptions. Configuation data can be provided for individual CPUs and individual boards.

    • Note. The structure of the config directory is proving overly complex, yet inflexible. It is likely it will be flattened in a future release of the scripts.

  • doc: The user manual for Embench.

  • src: The source for the benchmarks, one directory per benchmark.

  • support: The generic wrapper code for benchmarks, including substitutes for some library and emulation functions.

  • pylib: Support code for the python scripts

Licensing

Embench is licensed under the GNU General Public License version 3 (GPL3). See the COPYING file for details. Some individual benchmarks are also available under different licenses. See the comments in the individual source files for details.

The code base is OpenChain compliant, with SPDX license identifiers provided throughout.