PolyBench compiler autotuning dataset

This repository contains the dataset we used for the paper "A Collaborative Filtering Approah for the Automatic Tuning of Compiler Optimisations" available at https://arxiv.org/abs/2005.04092 and https://dl.acm.org/doi/abs/10.1145/3372799.3394361, which has been published in the 21st ACM SIGPLAN/SIGBED International Conference on Languages, Compilers, and Tools for Embedded Systems (LCTES 2020).

If you use these data please cite our paper:

@inproceedings{cereda2020collaborative,
  title={A Collaborative Filtering Approach for the Automatic Tuning of Compiler Optimisations},
  author={Cereda, Stefano and Palermo, Gianluca and Cremonesi, Paolo and Doni, Stefano},
  booktitle={The 21st ACM SIGPLAN/SIGBED Conference on Languages, Compilers, and Tools for Embedded Systems},
  pages={15--25},
  year={2020}
}

Optimisations

We follow the approach used in https://github.com/amirjamez/COBAYN, and all the files have the same format used in COBAYN. We consider 7 binary GCC flags:

• -funsafe-math-optimisations
• -fno-guess-branch-probability
• -fno-ivopts
• -fno-tree-loop-optimise
• -fno-inline-functions
• -funroll-all-loops
• -O2 or -O3 For a total of 128 possible combinations.

Instance

To collect the PolyBench dataset, we use an Amazon EC2 a1.medium instance, which is equipped with a single ARMv8 gravitron processor and 2GB of ram, Ubuntu Server 18.04 and PolyBench 4.2.1.

Execution times

The exec_times.csv file contains the execution time obtained by each benchmark available in PolyBench when compiled with each combination of optimisations. The first column contains the benchmark name, columns 2-8 specify the enabled compiler optimisations: X means the optimisation was turned off. We then report the execution times obtained using the MINI_DATASET, SMALL_DATASET, STANDARD_DATASET and LARGE_DATASET as defined in PolyBench. The fifth execution time column is zero as we did not use the EXTRALARGE_DATASET. The last column contains the execution time. We timed the execution using the PolyBench timing and enabling the FIFO scheduler.

MICA metrics

We also collected MICA metrics (https://github.com/boegel/MICA) using Intel PIN 3.10. We report these values in micaTable.csv, each line contains the benchmark name, the dataset size (mini, small, standard) and the values of the MICA metrics