Experimental setup for Evaluating Stochastic Floating-Point Superoptimization with STOKE

Table of Contents

• Table of Contents
• Overview
• Running the Experiments
  • Prerequisites
  • Steps to run the experiments
• Adding additional experiments
• Project Structure
• Useful links

Overview

This repository serves as an artifact to my submission for the TUDelft CSE research project. In this project I researched the effectiveness of STOKE on optimizing floating-point programs.

Running the Experiments

Prerequisites

• A 64 bits Linux operating system (The original experiments were run on Ubuntu 20.04.4 LTS x86_64, but any 64 bit linux distribution should work)
• GCC version 12.1
• GNU Make
• Docker
• Bash

Steps to run the experiments

1. When running the experiments for the first time you must run ./startup.sh, this will setup the STOKE docker container and starts it.
   • If the STOKE docker container has been stopped (by either running ./shutdown.sh or rebooting the computer) one should run ./restart.sh instead.
   • If you would like to connect to the docker container you can run ./connect.sh.
2. Now exit the docker container usingexit and change your directory to ./mount by running cd mount.
3. In order to time a function you run ./time.sh [directory name] [amount of times to synthesize] [amount of times to test] This will generate a work directory in the function directory in which 4 CSV files exist that contain the runtime and the program output.

Adding additional experiments

In order to run new experiments it is recommended to copy over the contents from an old experiment and only tweak the values you want to tweak.

1. Make a new directory in /mount for your new experiment: mkdir /mount/[name_of_new_experiment].
   • Note try to prevent any spaces in the path since this could lead to unexpected behavior of Make.
2. Copy over the contents of /mount/example to /mount[name_of_new_experiment].
3. Be sure to remove the old work directory/directories and the old test_cases.tc file.
4. Replace the function inside the /mount[name_of_new_experiment]/main.c file to the new function you would like to test.
5. Change the function parameter in the /mount[name_of_new_experiment]/Makefile to the name of the function you would like to optimize

Project Structure

• mount: Every file/folder that is in this directory is visible to the docker container.
• mount/[experiment_name]: every experiment should be stored in a separate directory
• mount/[experiment_name]/extract.conf: The extract configuration used by STOKE to list parameters for extracting the binary.
• mount/[experiment_name]/synthesize.conf: The synthesize configuration used by STOKE to list parameters for the synthesizing step.
• mount/[experiment_name]/testcase.conf: The test-case configuration used by STOKE for parameters for the test-case generation.
• mount/[experiment_name]/Makefile: Contains settings for general compiler options such as the location of where to find GCC.

Useful links

CSE3000 Research Projects Repository

STOKE github

Stochastic Superoptimization -- ASPLOS 2013

Stochastic Optimization of Floating-Point Programs with Tunable Precision -- PLDI 2014