HPAC is comprised of two components, namely the core and the harness. The core implements the approximate programming model. It extends the Clang/LLVM compiler and provides runtime support. The harness facilitates easy exploration of the approximate design space.

Currently, HPAC supports:

• Perforation:
  • ini: Drop the first iterations of a for loop.
  • fini: Drop the last iterations of a for loop.
  • random: Drop randomly iterations of a for loop.
  • small: Drop 1 iteration every N iterations.
  • large: Drop N-1 iterations every N iterations.
• Memoization:
  • iACT: Apply approximate memoization based on the inputs of a code region.
  • TAF: Apply approximate memoization based on the outputs of a code region.

HPAC on a x86_64 system has been successfully built with the following dependencies: following free projects:

• Ninja 1.9.0
• CMAKE 3.14
• gcc 8.3.1
• python 3.7.6
• pandas 1.2.2
• numpy 1.20.1
• matplotlib 3.3.4
• seaborn-0.11.1
• yaml 5.4.1

To build the HPAC compiler and all the infrastructure please execute the following commands:

git clone [email protected]:koparasy/HPAC.git
cd HPAC
./setup.sh 'PREFIX' 'NUM THREADS' 

The 'PREFIX' argument defines where to install all the HPAC related binaries and executables. The 'NUM THREADS' parameter define how many threads should the installation use. The installation script performs the following actions:

1. Configures, builds and installs clang/LLVM/OpenMP including the approximation extensions.
2. Configures, builds and installs the approximation library.
3. Creates a file, called 'hpac_env.sh', at the root of the project which should always be sourced before using HPAC.
4. Fetches the original applications of our evaluation.
5. Patches the applications with our extensions.
6. Configures and compiler the extended approximated applications.

The installation process can take a considerable amount of time. In our tested system, the complete installation took ~5 hours.

1. To use the harness, change directory to the HPAC root directory and source the hpac_env.shfile. The file was created during the installation. If the file is already source this step can be skipped.

source hpac_env.sh

2. Change directory to the 'approx/approx_utilities/' directory, Then issue the following command:

cd approx/approx_utilities/
python ./benchmark.py -t 'APPROX SPACE' -i 'CONFIG FILE' -p 'APPLICATION NAME' -b -r N -n '1,2,4,8' 

This command will perform the space exploration for the application called "APPLICATION NAME". It will read application configuration parameters from the file 'CONFIG FILE'. It will configure the approximation space to be 'APPROX SPACE'. For, each exploration point it will execute with different number of threads. In this example, it will explore the threads '1,2,4,8' and for each thread number it will execute 'N' experiments ('N' is defined with the '-r' option.

The HPAC harness executes the experiments and measures the execution time and the quality loss for different approximation techniques and parameters. All the results are stored into a "*.pkl" files under the 'analysis_results/interm_results/' directory. To view the contents of "*.pkl" file please issue the command:

python analysis_scripts/read.py analysis_results/interm_results/'FILE NAME'.pkl

Replace the 'FILE NAME' with the respective output file.

The output of this command outputs, to stdout, the contents of the table with the following columns.

The values of each row will correspond to the measured values of a specific experiment. The file is stored in Pandas format, and thus any data analysis technique can be applied easily on top of the data.

To run the analysis for each of the benchmarks please execute the following commands:

The scripts use two parameters. The first parameter defines the size of the exploration. There are two options, namely 'large' and 'small'. The final results are obtained using the 'large' option. However, the 'large' experiments may take days to finish. Therefore, there is a 'small' option which performs less experiments. The final outputs should not change significantly. Finally, the second option is visualize, which can take the value of 0 (disable visualization) or 1 (enable visualization). In the latter case, the visual outputs of the analysis are stored under the directory 'sc_paper_plots'. For example, if you execute:

./examples/blackscholes.sh 'small' 1

A figure called 'blackscholes_figure2.pdf' should be stored under the 'sc_paper_plots'. This figure is similar to Figure 2 of the paper submission.

To contribute to this repo please send a pull request on the develop branch of this repo.

This code was created by Konstantinos Parasyris ([email protected]) and Giorgis Georgakoudis, ([email protected]), assisted with design input from Harshitha Menon ([email protected]).

This repo is distributed under the terms of the Apache License (Version 2.0) with LLVM exceptions. Other software that is part of this repository may be under a different license, documented by the file LICENSE in its sub-directory.

All new contributions to this repo must be under the Apache License (Version 2.0) with LLVM exceptions.

See files LICENSE and NOTICE for more information.

SPDX License Identifier: "Apache-2.0 WITH LLVM-exception"

LLNL-CODE- 825539