ExaTron.jl implements a trust-region Newton solver for batched nonlinear programming on GPUs. Problems in the batch are solved in parallel by employing multiple thread blocks on GPUs. Our basic algorithm to solve each problem on GPUs is an extension of the algorithm by Lin and More and its code TRON.
Please check out this branch to see how to generate PTX code for our kernel and reproduce the experimental results in our technical report.
pkg> add ExaTron
@misc{ExaTron.jl.0.0.0,
author = {Kim, Youngdae and Pacaud, Fran\ccois and Kim, Kibaek},
title = {{ExaTron.jl: GPU-capable TRON solver in Julia}},
month = Mar,
year = 2021,
version = {0.0.0},
url = {https://github.com/exanauts/ExaTron.jl}
}
This material is based upon work supported by the U.S. Department of Energy, Office of Science, under contract number DE-AC02-06CH11357.