Spector
Spector is an OpenCL benchmark suite for FPGA. The particularity of these benchmarks is that each of them is outfitted with a range of optimization parameters (or knobs). These knobs can be tuned to create hundreds of unique designs per benchmark.
Each unique design can be compiled and run on FPGA to create design spaces that can be analyzed. We provide the performance and utilization characteristics of thousands of designs to encourage research and improve design space exploration.
Source codes
We provide the following algorithms as OpenCL benchmarks for FPGA:
- Breadth-First Search (graph traversal)
- Discrete Cosine Transform
- Time-Domain FIR filter
- Histogram calculation
- Merge sort
- Matrix multiplication
- Normal estimation
- Sobel filter
- Sparse matrix-vector multiplication
Compatibility
- The source codes have been tested under Linux, Ubuntu 14.04.
- The OpenCL codes have been tested on a Terasic DE5 board, compiled with the Altera OpenCL SDK v14.1.
- The scripts to generate the architectures have been tested on Ubuntu 14.04. Note that some of these scripts rely on soft links, which may not be available in some OS.
Design space data
We compiled thousands of unique designs with the Altera OpenCL SDK, executed them on a Terasic DE5 board, and recorded the performance and area results. These data are available in the results_and_analysis directory in MATLAB format.
Paper
Spector: An OpenCL FPGA Benchmark Suite
Quentin Gautier, Alric Althoff, Pingfan Meng and Ryan Kastner
@INPROCEEDINGS{Gautier2016,
author={Q. Gautier and A. Althoff and Pingfan Meng and R. Kastner},
booktitle={2016 International Conference on Field-Programmable Technology (FPT)},
title={Spector: An OpenCL FPGA benchmark suite},
year={2016},
pages={141-148},
keywords={field programmable gate arrays;high level synthesis;optimisation;statistical analysis;Spector;FPGA benchmark suite;high-level synthesis tools;design optimization;Altera OpenCL SDK;Terasic DE5 board;statistical analysis;Manuals;Benchmark testing;Discrete cosine transforms;Histograms;Open source software;Pipelines;Computational modeling},
doi={10.1109/FPT.2016.7929519},
month={Dec},}
License information
This work is released under the BSD 3 license provided in the LICENSE file in this folder.
Exceptions are:
- The OpenCL code of the BFS benchmark is released under the LGPL v2.1 license.
- The OpenCL code and most of the C/C++ code of the SPMV benchmark is released under the LGPL v2.1 license.
- Any unmodified code from an external source retains its original license.
Acknowledgments
This work was supported in part by an Amazon Web Services Research Education grant.