Mar 02

Print this Post

The Hack is on in 2018!

It’s that time of the year again! Developers of scientific software projects of any discipline come together with GPU experts from across the globe in Dresden to bring their science to the GPU! Stay tuned on this blog or by following the twitter hashtag #GPUHackDD.

So without further ado, here are this year's teams:

Team ZerialLab

We are after microscopy-based image analysis by solving a system of non-linear equations in multidimensional space (~100 dimensions) using an iterative convergence method. This system is solved for each pixel/voxel of a 2D/3D image. The home-made implementation gives similar calculation time on CPU (Intel I7, 12 threads/6 core) and high- end GPU (Kepler K20). We are doing software development for quantitative microscopy within biological laboratory for more than 15 years. The basic programming language is C++. Basic platforms are Windows for user interface and Linux for high-throughput computation on HPC systems. We are implementing some parts of our algorithms on GPU (OpenCL) from 2011.

Team Remeisen

Tessim is part of the Sixte software package developed at the Erlangen Centre for Astroparticle Physics (ECAP) (Sixte Homepage). Using the software package astronomers all around the world are able to perform simulations of X-ray observations as they would be carried out by a variety of different satellites in space. This makes it possible to determine the feasibility of observations, create analysis software in advance of satellite missions and to carry out performance studies of instruments in development. The Tessim tool itself consists of about 8000 LOC and it is currently implemented in C and uses standard libraries in addition to the GSL. The main user community is the consortium for the XIFU instrument. The satellite will be launched in 2028, our software package is used in all areas of the performance studies for the mission. A (to be developed) version using the GPU will be used mainly by the XIFU systems team.

Team Uni Graz

Our application is an Eikonal equation solver written in CUDA to solve the eikonal equation and simulate the wave propagation on the heart mesh. It is also used for the inverse problem in order to determining the patient specific cardiac conductivity parameters, where the solver is called by the optimization method many times until the optimization method finally converges. For this reason it is very important to have a very fast solver. We have implemented it in CUDA and currently it runs only in one GPU. We want to port it to cluster computing and the benefits are discussed above. LOC = 4248. It has the LGPL v2.1 license. We are part of a large community who are using the Eikonal equation for different research purposes including here the Medical and Technical Uni of Graz.

Team LeMonADErs

Calculating the interactions of macromolecules/soft matter is computationally demanding and recent simulations on CPU are limited in physical time and length scales. In our approach, we use the Bond Fluctuation Model (BFM)[1,2] as well-established Monte Carlo method for simulating coarse- grained polymeric materials. Our team has already developed and published a single-CPU-based open-source project "LeMonADE" on gitHub for the usage in the scientific community. We already ported the sequential BFM-algorithm onto the GPU with CUDA for parallel execution of the monomer trial and error procedure. Within the hackathon, we want to address further optimization of this approach and novel algorithms in the problem domain.

Team McStas-McXtrace

The McStas neutron Monte Carlo ray-tracing simulation package is a versatile tool for producing accurate simulations of neutron scattering experiments at reactors, short- and long-pulsed spallation sources. McXtrace is the X-ray counterpart, for X-ray scattering experiments at lab-, synchrotron- and free-electron laser sources. McStas and McXtrace are extensively used for design and optimization of instruments, virtual experiments, data analysis and user training. McStas was founded as an scientific, open-source collaborative code in 1997, McXtrace was founded in 2009. Technology wise, a user-written input file defined in our own DSL (lex+yacc grammar) provides input to a code-generator, and an ISO C code is generated, compiled and run. To (always) be able to simulate the (virtual) experiments faster than they are performed in reality, allow for in-experiment "data analysis", shorten execution time for "complicated" simulations, shorten development time for the instruments being constructed.

Team TU Dresden HF

Our application is a FDTD (finite difference time domain) solver for Maxwell's equation. This can be used to solve electromagnetic propagation problems. The current code base is about 8000 lines of code. The implementation is done in Python and has MPI support. However, the software currently runs on CPU only. Our goal is to port the datastructures and FDTD kernel to the GPU to increase the computational speed of the finite difference algorithms. The software is open-source with a BSD license.

Team Neuro Nerds

We have two applications in total: The first is about speeding up analysis of neurothermic images in neurosurgery. The second is about speeding up recurrent neural networks learning for deep learning in sensor applications. We think that there would be a performance gain of some hours or even days in the RNN case so our networks could become larger to learn more features from our data.


We have developed an in-house software for multi-resolution particle and mesh simulations. The novelty of our works comes from the new method : Adaptive particle representation (APR) we devised for details estimation and tracking for adaptive simulations. The APR was original developed for efficient and adaptive representation and storage of light sheet microscopy images, the method has now been extended for adaptive simulations of PDE's which reduces the number of DOF and the computational cost associated with it. We think, the GPU capabilities would massively decrease our compute times. More interestingly, we would like to quantify the performance characteristics of our novel algorithm. Potential applications include real- time adaptive simulations for system biology and computer graphics. We also envision an efficient and adaptive pipeline for image based system biology where fast and accurate simulations can be run in surfaces/geometries reconstructed from pixel data.

Team Ptycho_imaging

Phase reconstruction from intensity measurements obtained by ptychographical imaging. Ptychography is used to reconstruct 3D xray far field images from 2D scattering images. The algoritm is not a specific solution but a general purpose technique with a lot of applications. We use MATLAB with the Image Processing Toolbox running on laptop. We would like to run the code on the GPU.

Permanent link to this article: https://gcoe-dresden.de/the-hack-is-on-in-2018/

1 ping

  1. Meet the #GPUhackDD 2018 mentors » Dresden GPU Center of Excellence

    […] « The Hack is on in 2018! […]

Leave a Reply

Your email address will not be published. Required fields are marked *

You may use these HTML tags and attributes: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <s> <strike> <strong>