GPU Day 2019

Head to the Exascale …

Tibor Temesi¹

¹ Silicon Computers Ltd., Budapest, Hungary

HPC and AI can dramatically improve the way we live, work, and innovate. Based on a business economic study, enterprises within varied industries can realize significant revenue and profit by investing in HPC. There are several new technological challenges to meet growing performance demand. These requirements encourage HPC industry to make further significant technological improvements to achieve ten times and more higher performance systems than the current leaders at the top500.

The Exascale performance requires more processing elements, much bigger footprint, longer interconnects and on top it goes hand in hand with higher energy consumption. Among the others the cooling system and the cooling efficiency got into the focus because the current standard rack architecture is not working anymore in Exascale size, a new and very high-dense architecture is under development.

European supercomputing is still far behind the world, but the EuroHPC inspires to develop a World Class Supercomputing Ecosystem in Europe.

Contact address: [email protected]

Turning software into computer chips – Hastlayer

Zoltán Lehóczky¹

¹ Lombiq Ltd., Budapest, Hungary

Software is flexible, specialized hardware is extremely fast. So why not write software, then turn it into a computer chip? This is what Hastlayer (https://hastlayer.com) does by transforming .NET software into electronic circuits. The result is faster and uses less power while you simply keep on writing software. You may not be able to tell just by looking at it but behind some function calls now actually embedded hardware is working! (You wonder how? Check out what FPGAs are!) In this demo-packed session we'll get an overview of what Hastlayer is, why it is useful for developers like you and how to write Hastlayer-compatible software.

The future direction of SYCL and C++ Heterogeneous Programming

Michael Wong¹

¹ Codeplay Ltd., Edinburgh, United Kingdom

SYCL leads ISO C++ by supporting heterogeneous programming with modern C++. But when will ISO C++ catchup?

I would argue that heterogeneous support has been appearing in C++ since C++11, in bits and pieces. While there is no single TS or project that is named Heterogeneous, there is a quiet revolution to add support and I will highlight the C++ Parallel and Concurrency features that have driven in this direction.

In next week’s C++ Standard meeting, C++ will soon ratify C++20 with the release of the Committee Draft. It will undoubtedly be the largest C++release since C++11, with Concepts, Modules, Coroutines, Contracts all being major features. So,is there a deliberate future direction for C++? Or is it a random collection of features. I like to think there is with the establishment of a Direction Group.

SYCL is also expanding with vastly increased memberships, and many implementationsof SYCL. The latest being the open source clang upstream led by Intel but collaborated across several companies, similar to the OpenMP effort I started many years ago to upstream clang and support offloading. It is a good time to look forward to see SYCL evolvingwith future C++ directions.

Supercomputing on-demand

Gábor Varga¹

¹Microsoft Hungary Ltd., Budapest, Hungary

I attempt to clarify several myths associated with cloud in the context of high-performance computing. First of all, the best cloud is a hybrid cloud, with the ability to build friendly coexistence with on-premises (locally deployed) computing resources. I will discuss how cloud fares with key factors of success like security, cost, performance, innovation potential, regulatory compliance. I will talk about how several platforms ranging from CPU, GPU, FPGA (field programmable gate array) and even Cray can work together in a consistent manner in the same cloud. I will present a consistent AI environment built from open source tooling like Spark, containers, Hadoop, python and many other domain specific tools.

Report and plans on GPU accelerated HPC-s in Hungary

Zoltán Kiss¹

¹Governmental Information-Technology Development Agency (KIFÜ), Budapest, Hungary

KIFÜ operates the largest supercomputers in Hungary, serving the HPC needs of all kind of Hungarian research projects since 2001. Currently 8 running supercomputers with 8888 CPU cores and more than 250 GPUs provide nearly 450 Tflop/s total computing performance for the benefit of the users of universities and academic research institutes. In line with the international trends, more and more portion of our computing capacity is provided by GPGPUs. The presentation briefly introduces the HPC infrastructure and services of KIFÜ and gives information about how users can get access to the supercomputers focusing experiences with GPU usage. This talk will introduce current HPC, cloud and storage offerings, report on usage, and results of survey related to AI specific workloads. Another part will discuss the plans of deploying a 15PF HPC in Hungary.

Random Number Generation on GPUs

TBA¹

¹StreamHPC Ltd. Amsterdam, Netherlands

When a developer needs a lot of random numbers, he/she looks for support in the used language and uses it with default settings - after "seeding" it of course, which was mentioned somewhere. Ask any developer the difference between a real, pseudo and quasi random number and you get a blank face with a few exceptions. In this talk we discuss what random numbers are, what are the differences between popular generators like XORWOW, Mersenne Twister, Philox and Sobol, and how we created and optimized AMD's RNG GPU library. As we worked with both Nvidia and AMD GPUs, we'll also explain some key differences we encountered while developing the library. The goal of this talk is that you never look the same at a random number again and get insights in how modern AMD GPUs compare to Nvidia GPUs.

High-Performance Implementation Techniques of CUDA-based 1D and 2D Particle-in-Cell/MCC Plasma Simulations

Zoltán Juhász¹ , Péter Hartmann² , Zoltán Donkó²

¹ Department of Electrical Engineering and Information Systems, University of Pannonia, Veszprém, Hungary

² Department of Complex Fluids, Institute for Solid State Physics and Optics, Wigner Research Centre for Physics, Hungarian Academy of Sciences, Budapest, Hungary

The goal of our work is to create high-performance GPU implementations for Particle-in-Cell plasma simulations. The motivation lies in the fact that the simulation of low-pressure electrostatic plasma systems with Monte-Carlo collisions results in extreme execution times (ranging from several days to months). Speeding up simulations with parallelism using GPUs seemed to be the most promising implementation option. Over the past two years we have developed CUDA GPU implementations for one and two-dimensional plasma systems achieving two orders of magnitude speedup. Results of this work (approach to parallelization, achieved accuracy and performance) were already presented in the GPU Day 2017 and 2018 events.

In this talk, we would like to concentrate on the performance optimization aspects of this work. We describe the approach we took to port the original sequential implementation to GPU and in a step-by-step refinement fashion we demonstrate the performance benefits of various optimization methods, such as algorithmic modifications, CUDA kernel fusion, the use of shared memory, memory access optimizations, concurrent kernels and streams, high-performance atomic and warp operations.

Finally, we outline possible future improvement opportunities, including the use of multi-GPU systems.

Contact address: [email protected]

Purely Functional GPU Programming with Futhark

Troels Henriksen¹

¹ Department of Computer Science, University of Copenhagen, Copenhagen, Denmark

I present a pure functional array language, Futhark, along with its optimizing GPU-targeting compiler. I will focus on the language tradeoffs necessary to ensure the ability to efficiently generate high-performance GPU code from a high-level parallel language, as well as the compiler optimizations done to obtain good performance. I also demonstrate (nested) data-parallel array programming, a programming paradigm that enables concise programming of massively parallel systems. I show how Futhark code can be easily integrated with larger applications written in other language.

Modeling the effects of data locality

András Leitereg¹ , Dániel Berényi²

¹ Department of Programming Languages and Compilers, Faculty of Informatics, Eötvös Loránd University, Budapest, Hungary

² Wigner GPU Laboratory, Wigner Research Centre for Physics, Hungarian Academy of Sciences, Budapest, Hungary

Linear algebraic operations are vital in various areas of scientific computing, thus optimizing them is crucial. One aspect of these optimizations is focusing on data locality, that is, utilizing the hierarchical memory layout of the given computing hardware. In the case of linear algebra this is related to the order of operations, indices and access patterns. On different hardware different optimizations give the best performance and these details are usually captured in cost models. We are exploring an alternative way of learning this information with a recursive neural network (Tree Net) trained on benchmarking sample expressions.

Optimal scheduling in a Multi-GPU environment

Balázs Teréki¹

¹ AImotive Ltd., Budapest, Hungary

Developing a simulator for testing self-driving software raises many difficulties, including the real-time simulation of camera sensors. Simulating camera sensors in aiSim has two main tasks:

    1. the result must be as photorealistic as possible,

    2. the rendered images must be transferred from the GPU memory to the system memory

The architecture developed uses available hardware optimally through CPU and GPU scheduling. The presentation will focus on the problems we faced, and the solutions found while developing the GPU scheduler.

GPU-based real-time trajectory estimation from videos of vehicle-mounted cameras

Balázs Piszkor¹, István Gergő Gál¹, Tekla Tóth¹, Levente Hajder¹

¹ Department of Algorithms and their Applications, Faculty of Informatics, Eötvös Loránd University, Budapest, Hungary

Autonomous driving is one of the most popular research areas nowadays. Real-time sensing and control of vehicles require very high computational demand that cannot be possible without the usage of Graphics Processing Units (GPUs). The aim of our presentation is to show that real-time camera-based visual odometry is possible. Visual odometry is a basic problem for both computer vision and robotics.

The main idea behind the paper is that affine transformations between corresponding regions of stereo images give very useful information for the computation of camera motion. In the literature, only feature locations are used and spatial coordinates are computed by the so-called triangulation technique. We show that surface normals can also be computed from the affine transformations between corresponding image patches. Moreover, affine transformations depend on camera parameters as well, therefore camera extrinsic parameters, i.e. the relative pose between a stereo image pair, can also be retrieved.

For planar motion, which is common for vehicles in roads, only one affine transformation is required to estimate the extrinsic motion parameters. To the best of our knowledge, it is the first method in the literature that needs only a single correspondence to compute the relative camera pose between two images. Therefore, only local information has to be processed contrary to the mainstream pose estimators where at least two points correspondences are required even for planar motion.

The main challenge for a real trajectory estimator is that input data are highly contaminated by outliers, i.e. wrong point and/or affine correspondences. The well-know RANSAC algorithm is frequently applied to filter out outliers, however, CPU-implementations of RANSAC is very slow. In our case, histogram voting can also be used for robustification. Fortunately, both RANSAC and histogram voting can be efficiently parallelized, therefore the application of GPUs is preferred.

We have developed a CUDA implementation for the planar odometry problem. The most complex mathematical problem for the trajectory computation is to find the roots of a quartic polynomial. Closed-form solutions exist for this problem that can be implemented for GPU threads. The speed of the application is approximately 5FPS that is satisfactory for real-time vision of an autonomous vehicle.

Determinism and Low-Latency GPU Scheduling in OpenCL

Balázs Keszthelyi¹

¹ V-Nova Ltd., London, United Kingdom

Given today's successful GPU compute applications are primarily focusing on high-throughput massively sized workloads, latency and determinism requirements are things that are mostly addressable via sacrificing compute headroom. In this case-study, I am going to present some of the common pitfalls of trying to tame a throughput-oriented GPU architecture in a way to deliver deterministic output performance on a number of parallel processing threads, and also some techniques relying only merely on standard OpenCL 1.x features.

Gravitational Wave Data Analysis Using Naked OpenCL

Máté Ferenc Nagy-Egri¹

¹ Wigner GPU Laboratory, Wigner Research Centre for Physics, Hungarian Academy of Sciences, Budapest, Hungary

A recent study was conducted on a data analysis pipeline of narrow-band, all-sky search algorithm targeting continuous gravitational wave sources. Our hypothesis was that the employed algorithm is bandwidth limited, moreover the initial choice of floating-point precision being far too generous; hence reducing precision on various parts of the pipeline could be done without losing much precision on the results. Meanwhile, the algorithm was also ported to using OpenCL in hopes of broadening our set of possible target architectures. This talk summarizes our findings, the confirmation of our hypothesis, and experiences in implementing a generic OpenCL pipeline as well as peaks into the future, aiming to making our code more robust by leveraging SYCL.

GPU testing: past, present and the future

Ádám István Szűcs¹, Attila Kovács²

¹ Department of Computer Algebra, Faculty of Informatics, Eötvös Loránd University, Budapest, Hungary

² Mediso Ltd., Budapest, Hungary

Visual fidelity and accurate computations became even more important with the new arrival of ray tracing techniques. Projects involving realistic graphics and analytics are chasing real world experiences with simulation techniques on the boundary of real-time applications. Most of these technologies are aided with GPUs using the latest graphics and perform API operations with Vulkan.

With the growth rate in applications physically based rendering (PBR) and simulations, development time has increased, challenging the programmers and testing personnel. Most of these projects are based on shorter development and design times and often ends in buggy final products.

Our presentation will cover the results and findings of a pilot project implemented in a startup (Zeno) approaching physically based rendering with photon mapping and developing a GPU shader unit testing and pipeline reflection tool (FrameGraph) architecture for enhancing the development of rendering in real-time simulations [1]. The solution was implemented with the extension of SPIRV-X aided by custom automated toolchain steps to produce graph modules and equivalent shader conversions. Summing up our experiences the generation of explicit API calls and reflection on shader assets with testing almost always can lower the development time in the long run whereas it can assure great quality in visual fidelity.

We present the outline of our solution prototype which executes unit tests on the GPU by harnessing the Vulkan Validation Layer.

[1] Improving graphics programming with shader tests. April 2019, Pollack Periodica 14(1):35-46, DOI:10.1556/606.2019.14.1.4

Light-Field 3D Videoconferencing

Áron Cserkaszky¹, Attila Barsi¹, Zsolt Nagy¹, Gábor Puhr¹, Tibor Balogh¹, Péter András Kara¹

¹ Holografika Ltd., Budapest, Hungary

Light-field technology is often looked at as the final frontier of glasses-free 3D visualization, as no additional viewing gear is required to experience its capabilities to their full extent. Among the numerous industrial and commercial use cases, light-field telepresence stands out, as such natural visualization may significantly boost the sense of presence. In this paper, we present a fully-implemented real-time light-field 3D telepresence system. We provide a comprehensive analysis of the implementation of the one-way system, highlighting how the achieved capabilities satisfy the reasonable requirements towards such system. The paper also discusses future enhancements to the 3D telepresence system, since its true potential is yet to be fulfilled.

Contact address: [email protected]

Getting started with Vulkan

Viktor Makki¹

¹ Faculty of Electrical Engineering and Informatics, Budapest University of Technology and Economics, Budapest, Hungary

I began to work with the Vulkan API a year ago. I would like to share some tips and tricks I wish I knew before I got started. Many people I spoke with were afraid of Vulkan’s use, because of its hard initialization. So, I decided to write a library with configurable default settings to help the quicker initialization. I will introduce the Vulkan API and the prototype of this self-developed library, which simplifies Vulkan’s initialization, but keeps its configurability. First, I will speak about Vulkan’s architecture, how can we create an application and what do we need for drawing our first triangle. Next, I will present my library, how can we create classes that represent Vulkan objects with configurations different from the default. Then I will present the application side interface, how can we use the renderer library through this interface, which hides the implementation of the renderer layer. Finally, I will introduce the RenderDoc monitoring application to help debugging.

Machine learning in sciences

István Csabai¹

¹ Department of Physics of Complex Systems, Faculty of Science, Eötvös Loránd University, Budapest, Hungary

Deep learning has a deep relation with sciences: on one hand its motivation comes from neurobiology and also the math behind the algorithms is closely related to and based on the description of systems studied by physicists. On the other hand, the developed powerful machine learning tools can help not only everyday life like car driving or foreign language translation but they can bring new insights into basic research questions, too. I will review the role of machine learning in sciences and show some recent examples.

Interdisciplinary machine learning projects at FIAS

Olena Linnyk^1,2, K. Zhou¹, J. Steinheimer¹, N. Davis³, A. Rybicki³, W. Brylinski⁴, M. Gazdzincki^5,6, J. Kohl², I. Teetz², H. Stöcker^1,5,7

¹ Frankfurt Institute for Advanced Studies, Frankfurt am Main, Germany
² milch&zucker, Giessen, Germany
³ The Henryk Niewodniczanski Institute of Nuclear Physics, Polish Academy of Sciences, Cracow, Poland
⁴ Warsaw University of Technology, Warsaw, Poland
⁵ Goethe University Frankfurt, Frankfurt am Main, Germany
⁶ Jan Kochanowski University, Kielce, Poland
⁷ GSI Helmholtz Centre for Heavy Ion Research, Darmstadt, Germany

We present some of the interdisciplinary machine learning projects currently running at the Frankfurt Institute for Advanced Studies. The same machine learning methods based on the deep convolutional networks and Bayes analysis can be applied to the problems in the fields of theoretical physics, experimental detector design and social sciences. Online, real-time calibration is the expected next big improvements in the operation and design of the large sensor systems in science, research and industry alike. We are developing AI-algorithms to automatically and effectively calibrate detector components for experimental groups in the field of high energy physics. Previously, we have shown in theoretical simulations that deep neural networks can be trained to decode important underlying physics characteristics from the event-by-event distribution of the particles produced in the heavy-ion collisions. On the other hand, we are using the modern machine learning tools for the development of the AI-supported competence-oriented matching in human resources (project KIOMA). In this project we aim to fight the shortage of skilled workers by providing companies with data-based recruiting analytics and empowering the workers with market-driven self-education and career-planning strategies.

Dimensional causality

Ádám Ádám Zlatniczki¹, Dániel Fabó¹, András Sólyom¹, Loránd Erőss¹, András Telcs¹, Zoltán Somogyvári¹

¹Department of Computational Sciences, Wigner Research Centre for Physics, Hungarian Academy of Sciences, Budapest, Hungary

Causality is one, if not the most fundamental pillars of science. This lecture presents a new method, the Dimensional Causality (DC) method, devised to detect and quantify the probability of four types of causal relationships between two or more time series: independence, direct or circular causal connection, as well as the existence of a hidden common cause. This novel method is based on intrinsic dimension estimates of the joint and separate time delay embeddings of the time series. We validated our method on simulated examples and demonstrate its capabilities on human neuro-electrophysiological measurements. The beautiful information reduction method encapsulates two computational beasts. Data, the sample size is small relative to the dimension. Typical dimensions are 6-10 which would need at least 10⁶-10¹⁰ data points which is in scarcity in practice. The parameter space is also a nightmare, there are five parameters for the method but almost no preliminary information how to choose them.

Incentivizing exploration in curiosity-driven deep reinforcement learning

Patrik Reizinger¹, Márton Szemenyei¹

¹ Department of Control Engineering and Informatics, Budapest University of Technology and Economics, Budapest, Hungary

In deep reinforcement learning, the problem of credit assignment is a topic of active research, especially when the reward provided by the environment (called the external reward) is sparse – which is the case for the most environments modelling real tasks. Several model-based methods exist to mitigate the problem, among which curiosity-based ones utilize a so-called intrinsic reward (which is given by the agent to itself) to bridge the gap between external rewards. This approach usually rewards the agent if it is able to model the dynamics of the environment. A recent work of Pathak et al. does this by combining two factors: they define a loss for forward and one for backward dynamics, where the former penalizes if the next state cannot be predicted, while the latter does so for the action between two consecutive states. Their results are quite promising, as the agent trained with this definition of curiosity is able to learn general skills or to be less sensitive for noise injected into the states.

In this work, we investigate the problem of exploration for curiosity-based reinforcement learning agents. Our aim is to improve the exploration of the agent even if rewards are very sparse. Although the methodology referenced above implies that the agent will explore the state space, we argue that using an explicit incentive could help the agent to better explore. To achieve this, we add a new factor to the loss function of the A2C network of the model, which aims to provide a reward if states cannot be predicted well. This additional loss ensures that the agent will visit states never seen before (or which are hard to predict). Our approach is evaluated on standard tasks included in the Gym benchmark suit of OpenAI.

Contact address: [email protected]

Critical synchronization dynamics of the Kuramoto model on connectome and small world graphs

Géza Ódor¹

¹ Research Institute for Technical Physics and Materials Science, Centre for Energy Research, Hungarian Academy of Sciences, Budapest, Hungary

The hypothesis, that cortical dynamics operates near criticality also suggests, that it exhibits universal critical exponents which marks the Kuramoto equation, a fundamental model for synchronization, as a prime candidate for an underlying universal model. Here, we determined the synchronization behavior of this model by solving it numerically on a large, weighted human connectome network in an assumed homeostatic state. Since this graph has a topological dimension d < 4, a real synchronization phase transition is not possible in the thermodynamic limit, still we could locate a transition between partially synchronized and desynchronized states. At this crossover point we observe power-law-tailed synchronization durations, with 𝜏 ~ 1.2(1), away from experimental values for the brain. However, mimicking a network with inhibitory interactions by flipping the signs of the outgoing weights of a randomly selected 20% of nodes, we found 𝜏 ~ 1.9(2), which is in the range of human brain experiments. This implies that inhibitory interactions in brain models are necessary to describe experimental values. We provide results on a large two-dimensional lattice, having additional random, long-range links, finding a critical exponent 𝜏 ~1.6(1) for comparison.

Solving the Kuramoto Oscillator Model on Random Graphs

Jeffrey Kelling¹

¹ Helmholtz-Zentrum Dresden-Rossendorf, Department of Information Services and Computing, Dresden, Germany

The problem of synchronization is recently attracting much attention because it relates to current topics in science. The dynamics of electrical grids can be affected by de-synchronizations between supplier and consumer nodes. In brains, synchronization of neuronal activity plays an important role in most functions.

The Kuramoto model describes systems of coupled oscillators which, which exhibit non-trivial behavior on complex graphs, making it a suitable tool to study the synchronization dynamics of brains and other systems. Numerical solution of Kuramoto type ordinary differential equations for long times and large systems requires strong computation power, due to the inherent chaoticity of this nonlinear system.

In this talk we present a GPU implementation of Kuramoto models on sparse random graphs. The key to performance here, is the presented memory layout which supplements the SIMT usage of our design. Our implementation uses the VexCL [1] library to abstract GPUs and as a backend for ODE solvers provided by boost odeint [2].

[1] https://github.com/ddemidov/vexcl

[2] https://odeint.org

High-dimensional Hessian metric representation on GPGPUs

Bálint Daróczy¹

¹ Institute for Computer Science and Control, Hungarian Academy of Sciences, Budapest, Hungary

Optimization is one of the key problems of machine learning. The propagation graph (or “gradient graph”) of smooth hierarchical models, including neural networks, are based on relations of the compositional function in the tangent bundle. The complexity and the vector space describing the graph and the identified paths and non-trivial connections may result uncorrelated submanifolds while we can achieve efficient representations. Even though current GPGPUs and graph propagation frameworks are not ideal for high-dimensional, sparse representations, we show models achieving higher performance in image classification motivated by constraints of GPGPUs and recent results in graph factorization.

Detection of the bird song – a study on the collared flycatcher (Ficedula albicollis) with the help of deep neural networks

Sándor Zsebők¹, Máté Ferenc Nagy-Egri², Gergely Gábor Barnaföldi², Miklós Laczi¹ , Gergely Nagy¹, Éva Vaskuti¹, László Zsolt Garamszegi^1,3

¹ Department of Systematic Zoology and Ecology, Eötvös Loránd University, Budapest, Hungary

² Wigner GPU Laboratory, Wigner Research Centre for Physics, Hungarian Academy of Sciences, Budapest, Hungary

³ Institute of Ecology and Botany, Centre for Ecological Research, Hungarian Academy of Sciences, Vácrátót, Hungary

The animal bioacoustic investigations result enormous amount of digitized acoustic data, and we need effective automatic processing to extract the information content of the recordings. Our group is specialized for the song of collared flycatcher (Ficedula albicollis) and interested in the evolution of the animal acoustic signals. In the last 20 years, we obtained hundreds of hours of bird song recordings collected in natural environment, and we are seeking for automatic processing solutions. In this study, we chose an open-source, deep-learning image detection system (YOLO) to (1) find the specie’s specific songs of the collared flycatcher on the recordings and (2) to detect the small, discrete elements within the song. For the two tasks, we transformed the acoustic data into spectrogram images, and we trained two deep-learning models separately on our manually labelled and segmented database. The results are outstanding in both tasks and anticipating an order of magnitude less human effort in the acoustic processing than the manual method used before. Thanks to the new technique, we are able to address new biological questions that need large amount of acoustic data.

AI: from cats to medical imaging

Ákos Kovács¹

¹ Mediso Ltd., Budapest, Hungary

Mediso was not an AI company in the past years; it is a medical imaging company developing hardware and software of CT, MR, PET and SPECT modalities. Literature review showed us that neural networks can perform exceptionally well, but the articles focus on technology, and they do not detail real-life use cases about the application of these systems in industry or diagnostics. Apart from developing neural networks, which is not an easy task, the necessary steps to make them safely usable are much more difficult. We not only need to prove the validity of the algorithms, test the methods – as we have seen in the case of traditional software - we have to attack the networks to recognize their limits and capabilities. In nuclear medicine, some of the most important needs are having shorter examination time and lower amounts of radioisotope. Both of which have serious advantages, but result in decreased image statistics, higher noise levels. Sometimes the smallest features can imply serious medical conditions such as cancer. Therefore, it is not an acceptable option that an AI based noise filter dramatically improves the general image quality of the measurement, but reduces the detectability of a lesion.

We are focusing on artificial intelligence as much as about any new technology. We are using GPUs (CUDA, OpenCL) for reconstruction (e.g. Monte Carlo simulations) software, FPGA-s for signal processing.

Defining membrane boundaries of proteins using electron density maps – the MemBlob database and server

Georgina Csizmadia^1,#, Bianka Farkas^1,2,#, Eszter Katona^1,3, Gábor E. Tusnády⁴, Tamás Hegedűs¹

¹ Department of Biophysics and Radiation Biology, Semmelweis University, Budapest, Hungary and MTA-SE Molecular Biophysics Research Group, Hungarian Academy of Sciences, Budapest, Hungary

² Faculty of Information Technology and Bionics, Pázmány Péter Catholic University, Budapest, Hungary

³ University College London, London, United Kingdom

⁴ "Momentum" Membrane Protein Bioinformatics Research Group, Institute of Enzymology, Research Centre for Natural Sciences, Budapest, Hungary

# Equal contribution

Transmembrane (TM) proteins are highly significant drug targets, as they play an important role in many cellular processes. In order to understand TM protein function and to develop novel therapies targeting TM protein associated diseases, it is crucial to define the localization of transmembrane regions. While experimental data on the boundaries of transmembrane regions is scarce, cryo-electron microscopy (cryo-EM) density maps contain still unexploited information on the membrane embedment. To extract this information, we developed a computational pipeline, which requires a cryo-EM map, the corresponding atomistic structure, and the potential bilayer orientation determined by the TMDET algorithm as input. The resulted output includes the residues assigned to the bulk water phase, lipid interface, and the lipid hydrophobic core. We used this method to build a database, which contains the information on TM regions of protein structures based on the corresponding cryo-EM maps with a resolution better than 4 Å. We also created a web application to allow the analysis of unpublished densities. Unlike popular membrane predictors, our method presents the extracted membrane region as a volume with boundaries that follows the shape of the lipid environment and is not a slab with parallel planes. Utilizing our pipeline, we obtained the first data set at atomic resolution that allows the comparison of in silico predictions with experiments. The MemBlob database and server is available at http://memblob.hegelab.org.

Support: NKFIH-111678, NKFI-127961, CFF HEGEDU18I0, KIFÜ HPC, MTA Wigner GPU Laboratory, NVIDIA Corporation, Semmelweis Science and Innovation Fund

Discovering the chloride conducting pathway of the CFTR channel using in silico methods

Bianka Farkas^1,2, Hedvig Tordai¹, Rita Padányi^1,3, János Gera⁴, Gábor Paragi⁴, Tamás Hegedűs^1,3

¹ Department of Biophysics and Radiation Biology, Semmelweis University, Budapest, Hungary

² Faculty of Information Technology and Bionics, Pázmány Péter Catholic University, Budapest, Hungary

³ MTA-SE Molecular Biophysics Research Group, Hungarian Academy of Sciences, Budapest, Hungary

⁴ Department of Medical Chemistry, University of Szeged, Szeged, Hungary

Cystic fibrosis (CF) is a lethal monogenic disease caused by pathogenic variants of the CFTR/ABCC7 chloride channel, a member of the ATP Binding Cassette (ABC) protein superfamily. Most of the CF mutations affect protein folding and stability, leading to reduced apical anion conductance of epithelial cells. The recently published cryo-EM structures of full- length human and zebrafish CFTR provide a good model to gain insight into the relationship between structure and function of CFTR variants. While some of the structures were determined in the phosphorylated and ATP-bound active state, none of the static structures shows an open pathway for chloride permeation. Our goal is to characterize the CFTR structure and dynamics based on both experimental and homology models, using in silico methods. Therefore, we performed extensive molecular dynamics simulations to generate a conformational ensemble of the protein. Channel detecting algorithms were applied to identify conformations with open channel and the possible pathways were characterized. Since chloride ions entered the pathway in our equilibrium simulations without traversing the bottleneck region, we performed metadynamics simulations to describe the passage through this region. We defined the chloride pathway of CFTR at atomic resolution which suggests the existence of several intra- and extracellular entry sites. One of the extracellular exits includes hydrophobic lipid tails that may explain the lipid-dependency of CFTR function. In summary, our in silico study describes a potential chloride pathway based on a recent cryo-EM structure and it helps to understand the gating mechanism of the wild type and mutant CFTR proteins.

Support: NKFIH-111678, NKFI-127961, CFF HEGEDU18I0, KIFÜ HPC, MTA Wigner GPU Laboratory, NVIDIA Corporation, and Semmelweis Science and Innovation Fund.

Contact address: [email protected]

C++QED: a framework for simulating open quantum dynamics – the first ten years

András Vukics¹

¹ Department of Quantum Optics and Quantum Information, Institute for Solid State Physics and Optics, Wigner Research Centre for Physics, Hungarian Academy of Sciences, Budapest, Hungary

C++QED is a versatile open-source C++/Python application-programming framework for simulating open quantum dynamics. It allows users to build arbitrarily complex interacting quantum systems from elementary free subsystems and interactions and simulate their time evolution with a number of available time-evolution drivers. In my presentation, I will first sketch the outlines of C++QED, concentrating on two unique features:

    1. the fundamental design idea of the framework, which relies heavily on the multi-array concept and compile-time algorithms (C++ template metaprogramming) [1-4]

    2. a modification of the Monte Carlo wave-function method to use adaptive time step [1,5]

Subsequently, I will present some important examples of usage that occurred during the first decade of the history of the framework, some instructive limitations of the design, and the problems and challenges that C++QED has faced during this time and is facing nowadays.

References:

[1] cppqed.sf.net

[2] A. Vukics and H. Ritsch. C++QED: an object-oriented framework for wave-function simulations of cavity QED systems. Eur. Phys. J. D, 44:585–599, 2007.

[3] András Vukics. C++QEDv2: The multi-array concept and compile-time algorithms in the definition of composite quantum systems. Comp. Phys. Comm., 183:1381–1396, 2012.

[4] Raimar Sandner, András Vukics. C++QEDv2 Milestone 10: A C++/Python application-programming framework for simulating open quantum dynamics. Comp. Phys. Comm., 185:2380–2382, 2014.

[5] M. Kornyik, A. Vukics. The Monte Carlo wave-function method: A robust adaptive algorithm and a study in convergence. Comp. Phys. Comm., 238:88–101, 2019.

MPGOS: A modular and general-purpose program package to solve a large number of independent ODE systems

Ferenc Hegedűs¹

¹ Department of Hydrodynamic Systems, Budapest University of Technology and Economics, Budapest, Hungary

In many fields of science, the governing equations of a physical problem can be described by an ordinary differential equation system (ODEs). The application of the massively parallel environment of GPUs on such systems is not trivial, since an initial value problem of an ODE is serial in nature. However, if large number of parameters or initial conditions are involved during a research study, the overall problem can be decomposed into a large number of decoupled sub-problems where each parameter combination and/or initial conditions represent an individual and independent task. Therefore, the whole computational problem became suitable for professional GPGPUs using the SIMT (single instruction multiple thread) paradigm.

The main aim of the present study is to introduce a modular and general-purpose program package capable to handle the aforementioned problem. The name of the package is Massively Parallel GPU-ODE Solver (MPGOS) and it is free to use under an MIT license, for details see the official website www.gpuode.com. The parallelization strategy is very simple: a GPU thread is assigned to each system having different parameters and/or initial conditions (the individual systems themselves are independent). The program package is written in C++ and CUDA C. Event handling is incorporated to detect special points of a trajectory. As a special feature, the user can specify parameters shared among each system (e.g. a differentiation matrix originated from the discretization of a PDE), which are automatically loaded into the shared memory of the GPU. Moreover, user-programmable parameters can also be defined to store the special features of the solutions. They can be updated after every successful event detection or time steps by the user through pre-declared device functions (the control flow is absolutely under the control of the user). In this way, for instance, non-smooth dynamics can be efficiently handled. In the first version of the code, only explicit solvers are implemented: the 4th order explicit and adaptive Runge–Kutta–Cash–Karp method with 5th embedded error estimation, and the well-known 4th order explicit Runge–Kutta technique with fixed time step.

One of the main strengths of the package is the user friendly framework that automatically manages the allocation of the data structure and the data transfer between the CPU side and the GPU side. Thus, it can almost completely hide GPU programming form the user. In case of multi-GPU systems, the distribution of the task to different GPUs can also be easily done.

Chasing a quantum anisotropy with GPUs

Dénes Molnár^1,2, Chris H. Greene¹, Fuqiang Wang¹

¹ Purdue University, West Lafayette, Indiana, United States

² Department of Theoretical Physics, Wigner Research Centre for Physics, Hungarian Academy of Sciences, Budapest, Hungary

One of the most exciting observations in high-energy heavy-ion collisions is an "elliptical" (left-right vs top-down) momentum anisotropy 𝑣₂=⟨cos (2𝜙)⟩ in the plane transverse to the colliding beams. The standard interpretation is that 𝑣₂comes from a femtometer-scale asymmetric droplet of hot and dense plasma, with a roughly elliptical initial density profile, which droplet then expands hydrodynamically [see, e.g., Gale et al, arXiv:1301.5893 for a review]. On the other hand, it is a general feature of quantum mechanics that coordinate and momentum space are intimately related via the uncertainty relation. Thus, a nonzero eccentricity in coordinate space automatically implies a nonzero 𝑣₂. This intrinsic "quantum" anisotropy vanishes for very large systems, and it also vanishes in the classical limit because the spacing between energy levels becomes negligible compared to 𝑘𝑇. So the intrinsic 𝑣₂ is completely missed by the canonical formulation of statistical physics in terms of phase space integrals.

In an earlier work [arxiv:1404.4119v1] we showed that the quantum anisotropy is sizeable in Au+Au collisions at the Relativistic Heavy Ion Collider (RHIC) - in particular, we estimated for the pion 𝑣₂ tens of percents at 𝑝_𝑇 ~ 1-2 GeV transverse momentum. However, that calculation assumed nonrelativistic 𝐾=𝑝²/2𝑀 kinetic energy (questionable for 𝑝_𝑇/Mpi ~ 10). Very recently [arxiv:1404.4119v2] we have also computed the anisotropy, numerically, for massless particles (i.e., 𝐾 = |𝑝|), and found a rather different pion 𝑣₂(𝑝_𝑇)shape with percent-level magnitude only. This leaves the question wide open regarding how large the quantum 𝑣₂ is, quantitatively, for correct relativistic 𝐾 = √𝑝²+𝑚².

The calculation of the intrinsic anisotropy takes two steps: first the Hamiltonian of the system is diagonalized in a large computational basis (N x N matrix with N ~ 10 5 ), and then the squares of the eigenfunctions are summed using thermal weights to construct 𝑣₂. With relativistic kinetic energy, the computation of the matrix elements is the most time-consuming task. We implement this part of the calculation on GPUs, and present preliminary pion 𝑣₂(𝑝_𝑇) results with unapproximated kinetic energy.