This repository is a FAIR-compliant hub for GPU auto-tuning benchmarking resources. It provides fully brute-forced search spaces (in original and T4 format) and their source files (kernel files and T1 input files). Anyone can contribute by providing new kernels and brute-forced search spaces on new configurations, please do so!
The brute-forced files can be used with the autotuning-methodology to compare optimization algorithms accross a wide variety of search spaces, with T1/T4 compliant auto-tuning frameworks, and in addition with Kernel Tuner for hyperparameter tuning optimization algorithms, without constant access to the original hardware. Please see "FAIR Sharing of Data in Autotuning Research (Vision Paper)" for more information on the T1 and T4 formats.
To automatically compress new cachefiles when committing and decompress new cachefiles when checking out, run git config --local core.hooksPath .githooks/.
To get the decompressed files after cloning, run the above command and checkout with git checkout main.
24 fully brute-forced search spaces are currently available, as a product of the following kernels and GPUs. For more information on the kernels and GPUs, see below.
Kernels:
- GEMM
- Convolution
- Hotspot
- Dedispersion
GPUs:
- Nvidia A100
- Nvidia A4000
- NVIDIA A6000
- AMD MI250X
- AMD W6600
- AMD W7800
Important: this is a live repository, subject to change. For persistent identifiers use the DOI of a specific release. If you have questions or suggestions please submit an issue.
kernelscontains for each kernel the contains the kernel files, T1 input format JSON file, and the script for the auto-tuning.cachefilescontains the brute-forced search spaces. Each kernel has its own folder, which contains both a T4 format output file and an original cache file for each GPU the kernel has been brute-forced on.utilitiescontains utility scripts.
Additional information on the benchmarks is provided here.
The four kernels are the dedispersion, convolution, hotspot, and GEMM kernels as used in "Bringing Auto-Tuning to HIP: Analysis of Tuning Impact and Difficulty on AMD and Nvidia GPUs", widely used in astronomy, image processing, material science, and linear algebra respectively. A brief description of each is provided:
- Dedispersion is a signal processing kernel that reconstructs radio signals distorted by interstellar dispersion by applying a range of dispersion measures to time-domain samples across multiple frequency channels.
- The 2D Convolution kernel performs image filtering by computing weighted sums over image regions.
- Hotspot is a thermal simulation kernel that estimates processor temperature by iteratively solving differential equations based on simulated power and initial temperature inputs, producing a temperature grid as output.
- GEMM (General Matrix-Matrix Multiplication) is a widely-used linear algebra operation implemented in CLBlast for large dense matrices.
The hardware used are the following: