Skip to content

Files

Failed to load latest commit information.

Latest commit

 Cannot retrieve latest commit at this time.

History

History

benchmarks

README.md

Benchmarks

This directory contains all the scripts and configuration files needed to reproduce the numerical results presented in the paper Recovering Sparse DFT from Missing Signals via Interior Point Method on GPU.

Overview

Each script in this folder benchmarks different aspects of our GPU-accelerated interior-point solver and FFT implementations, comparing them against CPU-based references.

Requirements

Ensure you have the appropriate hardware drivers installed (CUDA for NVIDIA GPUs, ROCm for AMD GPUs).

Installation

  1. Launch Julia with the project environment:
julia --project=.
  1. Instantiate the environment:
using Pkg
Pkg.instantiate()

Usage

To run a benchmark script, use one of the following commands:

julia --project=. -e 'include("benchmarks_cufft.jl")'
julia --project=. -e 'include("benchmarks_rocfft.jl")'
julia --project=. -e 'include("cpu_vs_gpu.jl")'
julia --project=. -e 'include("crystal.jl")'

Scripts

  • benchmarks_cufft.jl

Compares cuFFT (via CUDA.jl) against FFTW (via FFTW.jl) on problems of various sizes. Measures execution time for fft and ifft operations on random data.

  • benchmarks_rocfft.jl

Compares rocFFT (via AMDGPU.jl) against FFTW (via FFTW.jl). Similar to the cuFFT benchmarks; results were not included in the final paper.

  • cpu_vs_gpu.jl

Benchmarks our compressed sensing solver on CPU vs GPU across a range of problem sizes (artificial test cases).

  • crystal.jl

Applies the same solver to a real-world problem of 104 million variables, comparing CPU and GPU performance on a crystallographic dataset.

Preferences

To enable unified memory by default on the GH200, create a file named LocalPreferences.toml in this directory with the following content:

[CUDA]
default_memory = "unified"

Acknowledgments

We thank JLSE for providing access to the NVIDIA GH200 used in our experiments.