Matrix Computations on the GPU with Arrayfire for Python and C/C++

Total Page：16

File Type：pdf, Size：1020Kb

Matrix Computations on the GPU with ArrayFire for Python and C/C++ by Andrzej Chrzȩszczyk of Jan Kochanowski University Phone: +1.800.570.1941 • [email protected] • www.accelereyes.com The “AccelerEyes” logo is a trademark of AccelerEyes. Foreward by John Melonakos of AccelerEyes One of the biggest pleasures we experience at AccelerEyes is watching programmers develop awesome stuff with ArrayFire. Oftentimes, ArrayFire programmers contribute back to the community in the form of code, examples, or help on the community forums. This document is an example of an extraordinary contribution by an ArrayFire programmer, written entirely by Andrzej Chrzȩszczyk of Jan Kochanowski University. Readers of this document will find it to be a great resource in learning the ins-and-outs of ArrayFire. On behalf of the rest of the community, we thank you Andrzej for this marvelous contribution. Phone: +1.800.570.1941 • [email protected] • www.accelereyes.com The “AccelerEyes” logo is a trademark of AccelerEyes. Foreward by Andrzej Chrzȩszczyk of Jan Kochanowski University In recent years the Graphics Processing Units (GPUs) designed to efficiently manipulate computer graphics are more and more often used to General Purpose computing on GPU (GPGPU). NVIDIA’s CUDA and OpenCL platforms allow for general purpose parallel programming on modern graphics devices. Unfortunately many owners of powerful graphic cards are not experienced programmers and can find these platforms quite difficult. The purpose of this document is to make the first steps in using modern graphics cards to general purpose computations simpler. In the first two chapters we want to present the ArrayFire software library which in our opinion allows to start computations on GPU in the easiest way. The necessary software can be downloaded from: http://www.accelereyes.com/products/arrayfire In the present text we describe the ArrayFire 1.1 free version. It allows for efficient dense matrix computations in single precision on single GPU. The readers interested in double precision linear algebra, multiple GPUs, or sparse matrices should consider ArrayFire Pro. Phone: +1.800.570.1941 • [email protected] • www.accelereyes.com The “AccelerEyes” logo is a trademark of AccelerEyes. .

Copy Link

Recommended publications

Analysis of GPU-Libraries for Rapid Prototyping Database Operations
Analysis of GPU-Libraries for Rapid Prototyping Database Operations A look into library support for database operations Harish Kumar Harihara Subramanian Bala Gurumurthy Gabriel Campero Durand David Broneske Gunter Saake University of Magdeburg Magdeburg, Germany fi[email protected] Abstract—Using GPUs for query processing is still an ongoing Usually, library operators for GPUs are either written by research in the database community due to the increasing hardware experts [12] or are available out of the box by device heterogeneity of GPUs and their capabilities (e.g., their newest vendors [13]. Overall, we found more than 40 libraries for selling point: tensor cores). Hence, many researchers develop optimal operator implementations for a specific device generation GPUs each packing a set of operators commonly used in one involving tedious operator tuning by hand. On the other hand, or more domains. The benefits of those libraries is that they are there is a growing availability of GPU libraries that provide constantly updated and tested to support newer GPU versions optimized operators for manifold applications. However, the and their predefined interfaces offer high portability as well as question arises how mature these libraries are and whether they faster development time compared to handwritten operators. are fit to replace handwritten operator implementations not only w.r.t. implementation effort and portability, but also in terms of This makes them a perfect match for many commercial performance. database systems, which can rely on GPU libraries to imple- In this paper, we investigate various general-purpose libraries ment well performing database operators. Some example for that are both portable and easy to use for arbitrary GPUs such systems are: SQreamDB using Thrust [14], BlazingDB in order to test their production readiness on the example of using cuDF [15], Brytlyt using the Torch library [16].
[Show full text]
Conference Program
Ernest N. Morial ConventionConference Center Program • New Orleans, Louisiana HPC Everywhere, Everyday Conference Program The International Conference for High Performance Exhibition Dates: Computing, Networking, Storage and Analysis November 17-20, 2014 Sponsors: Conference Dates: November 16-21, 2014 Table of Contents 3 Welcome from the Chair 67 HPC Impact Showcase/ Emerging Technologies 4 SC14 Mobile App 82 HPC Interconnections 5 General Information 92 Keynote/Invited Talks 9 SCinet Contributors 106 Papers 11 Registration Pass Access 128 Posters 13 Maps 152 Tutorials 16 Daily Schedule 164 Visualization and Data Analytics 26 Award Talks/Award Presentations 168 Workshops 31 Birds of a Feather 178 SC15 Call for Participation 50 Doctoral Showcase 56 Exhibitor Forum Welcome 3 Welcome to SC14 HPC helps solve some of the SC is fundamentally a technical conference, and anyone world’s most complex problems. who has spent time on the show floor knows the SC Exhibits Innovations from our community have program provides a unique opportunity to interact with far-reaching impact in every area of the future of HPC. Far from being just a simple industry science —from the discovery of new exhibition, our research and industry booths showcase recent 67 HPC Impact Showcase/ drugs to precisely predicting the developments in our field, with a rich combination of research Emerging Technologies next superstorm—even investment labs, universities, and other organizations and vendors of all 82 HPC Interconnections banking! For more than two decades, the SC Conference has types of software, hardware, and services for HPC. been the place to build and share the innovations that are 92 Keynote/Invited Talks making these life-changing discoveries possible.
[Show full text]
“GPU in HEP: Online High Quality Trigger Processing”
“GPU in HEP: online high quality trigger processing” ISOTDAQ 15.1.2020 Valencia Gianluca Lamanna (Univ.Pisa & INFN) G.Lamanna – ISOTDAQ – 15/1/2020 Valencia TheWorldin2035 2 The problem in 2035 15/1/2020 Valencia 15/1/2020 – FCC (Future Circular Collider) is only an example Fixed target, Flavour factories, … the physics reach will be defined ISOTDAQ ISOTDAQ – by trigger! What the triggers will look like in 2035? 3 G.Lamanna The trigger in 2035… … will be similar to the current trigger… High reduction factor High efficiency for interesting events Fast decision High resolution …but will be also different… The higher background and Pile Up will limit the ability to trigger 15/1/2020 Valencia 15/1/2020 on interesting events – The primitives will be more complicated with respect today: tracks, clusters, rings ISOTDAQ ISOTDAQ – 4 G.Lamanna The trigger in 2035… Higher energy Resolution for high pt leptons → high-precision primitives High occupancy in forward region → better granularity Higher luminosity track-calo correlation Bunch crossing ID becomes challenging, pile up All of these effects go in the same direction 15/1/2020 Valencia 15/1/2020 More resolution & more granularity more data & more – processing What previously had to be done in hardware may ISOTDAQ ISOTDAQ – now be done in firmware; What was previously done in firmware may now be done in software! 5 G.Lamanna Classic trigger in the future? Is a traditional “pipelined” trigger possible? Yes and no Cost and dimension Getting all data in one place • New links -> data flow • No “slow”
[Show full text]
GPU-Accelerated Applications for HPC Industries| NVIDIA
GPU-ACCELERATED APPLICATIONS GPU‑ACCELERATED APPLICATIONS Accelerated computing has revolutionized a broad range of industries with over three hundred applications optimized for GPUs to help you accelerate your work. CONTENTS 01 Computational Finance 02 Data Science & Analytics 02 Defense and Intelligence 03 Deep Learning 03 Manufacturing: CAD and CAE COMPUTATIONAL FLUID DYNAMICS COMPUTATIONAL STRUCTURAL MECHANICS DESIGN AND VISUALIZATION ELECTRONIC DESIGN AUTOMATION 06 Media and Entertainment ANIMATION, MODELING AND RENDERING COLOR CORRECTION AND GRAIN MANAGEMENT COMPOSITING, FINISHING AND EFFECTS EDITING ENCODING AND DIGITAL DISTRIBUTION ON-AIR GRAPHICS ON-SET, REVIEW AND STEREO TOOLS WEATHER GRAPHICS 10 Oil and Gas 11 Research: Higher Education and Supercomputing COMPUTATIONAL CHEMISTRY AND BIOLOGY NUMERICAL ANALYTICS PHYSICS 16 Safety & Security Computational Finance APPLICATION DESCRIPTION SUPPORTED FEATURES MULTI-GPU SUPPORT Aaon Benfield Pathwise™ Specialized platform for real-time hedging, Spreadsheet-like modeling interfaces, Yes valuation, pricing and risk management Python-based scripting environment and Grid middleware Altimesh’s Hybridizer C# Multi-target C# framework for data parallel C# with translation to GPU or Multi-Core Yes computing. Xeon Elsen Accelerated Secure, accessible, and accelerated back- Web-like API with Native bindings for Yes Computing Engine (TM) testing, scenario analysis, risk analytics Python, R, Scala, C. Custom models and and real-time trading designed for easy data streams are easy to add. integration and rapid development. Global Valuation Esther In-memory risk analytics system for OTC High quality models not admitting closed Yes portfolios with a particular focus on XVA form solutions, efficient solvers based on metrics and balance sheet simulations. full matrix linear algebra powered by GPUs and Monte Carlo algorithms.
[Show full text]
John Melonakos, Arrayfire; James Reinders, Intel ______
Episode 20: Acceleration at the Edge Host: Nicole Huesman, Intel Guests: John Melonakos, ArrayFire; James Reinders, Intel ___________________________________________________________________________________ Nicole Huesman: Welcome to Code Together, an interview series exploring the possibilities of cross- architecture development with those who live it. I'm your host, Nicole Huesman. The industry is increasingly embracing open standards like modern C++ and SYCL to address the need for programming portability and performance across heterogeneous architectures. This is especially important with the explosion of data-centric workloads in areas like HPC, AI, and machine learning. John Melonakos joins us today. John is the CEO and co-founder of ArrayFire, a software development and consulting company with a passion for AI and GPU acceleration projects and expertise in machine learning and computer vision. Thanks so much for being here, John! John Melonakos: Thank you for having me. I look forward to the conversation. Nicole Huesman: And James Reinders also joins us. As an engineer at Intel, James brings deep expertise and parallel computing having authored over 10 technical books on the topic, including a recent book about Data Parallel C++, or DPC++. He focuses on promoting parallel programming in a heterogeneous world. Great to have you with us, James! James Reinders: Happy to be here. Nicole Huesman: This episode was really inspired when I saw a recent blog that John authored. It really caught my attention. While ArrayFire has expertise in CUDA and OpenCL, in this blog, John writes about the open SYCL standard and its implementations. John, I understand you're a champion of things that are open, programmable and maximize performance.
[Show full text]
Arrayfire Webinar: Opencl and CUDA Trade-Offs and Comparisons GPU Software Features Programmability Portability Scalability
ArrayFire Webinar: OpenCL and CUDA Trade-Offs and Comparisons GPU Software Features Programmability Portability Scalability Performance Community Leading GPU Computing Platforms Programmability Portability Scalability Performance Community Performance • Both CUDA and OpenCL are fast • Both can fully utilize the hardware • Devil in the details: – Hardware type, algorithm type, code quality Performance • ArrayFire results at end of webinar Scalability • Laptops --> Single GPU Machine – Both CUDA and OpenCL scale, no code change • Single GPU Machine --> Multi-GPU Machine – User managed, low-level synchronization • Multi-GPU Machine --> Cluster – MPI Scalability Interesting developments: • Memory Transfer Optimizations – CUDA GPUDirect technology • Mobile GPU Computing – OpenCL available on ARM, Imgtec, Freescale, … Scalability • Laptops --> Single GPU Machine – ArrayFire’s JIT optimizes for GPU type • Single GPU Machine --> Multi-GPU Machine – ArrayFire’s deviceset() function is super easy • Multi-GPU Machine --> Cluster – MPI Portability • CUDA is NVIDIA-only – Open source announcement – Does not provide CPU fallback • OpenCL is the open industry standard – Runs on AMD, Intel, and NVIDIA – Provides CPU fallback Portability • ArrayFire is fully portable – Same ArrayFire code runs on CUDA or OpenCL – Simply select the right library Community • NVIDIA CUDA Forums – 26,893 topics • AMD OpenCL Forums – 4,038 topics • Stackoverflow CUDA Tag – 1,709 tags • Stackoverflow OpenCL Tag – 564 tags Community • AccelerEyes GPU Forums – 1,435 topics –
[Show full text]
GPGPU IMPLEMENTATION of SCHRÖDINGER's SMOKE for UNITY3D Oleg Iakushkin A, Anastasia Iashnikova, Olga Sedova
Proceedings of the VIII International Conference "Distributed Computing and Grid-technologies in Science and Education" (GRID 2018), Dubna, Moscow region, Russia, September 10 - 14, 2018 GPGPU IMPLEMENTATION OF SCHRÖDINGER'S SMOKE FOR UNITY3D Oleg Iakushkin a, Anastasia Iashnikova, Olga Sedova Saint Petersburg State University, 7/9 Universitetskaya nab., St. Petersburg, 199034, Russia E-mail: a [email protected] The paper describes an algorithm for Eulerian simulation of incompressible fluids — Schrödinger's Smoke. The algorithm is based on representing models as a system of particles. Each particle represents a small portion of a fluid or amorphous material. A particle has a certain ‘lifespan’, during which it may undergo various changes. The Schrödinger's Smoke solver algorithm was implemented in Unity3D environment. We used ArrayFire library to transfer the bulk of computational load relating to simulation of physical processes to GPGPU — it allowed real-time interaction with the model. The solution we developed allows to model interactions between vortex rings — i.e., their collisions and overlapping — with a high degree of physical accuracy. Keywords: particle system, GPGPU, Unity3D, ArrayFire © 2018 Oleg O. Iakushkin, Anastasia P. Iashnikova, Olga S. Sedova 467 Proceedings of the VIII International Conference "Distributed Computing and Grid-technologies in Science and Education" (GRID 2018), Dubna, Moscow region, Russia, September 10 - 14, 2018 1. Introduction We examined the work on Schrödinger’s Smoke [1] that demonstrates a new approach to modelling Eulerian incompressible fluids. We transferred its algorithm to the multi-platform Unity3D development environment using ArrayFire, a library for transferring computing to GPGPU. 2. Approach Initially, we wanted to use the Compute Shader technology available in Unity3D as the main computational tool in order to transfer most of the algorithm calculations to the GPGPU.
[Show full text]
The Artistry of Molecular Animation >>
Cambridge Healthtech Media Group NEXT-GEN TECHNOLOGY • BIG DATA • PERSONALIZED MEDICINE JANUARY 2012 Data Visualization ALL NEW DIGITAL The Artistry EDITION of Molecular Animation >> FEATURE | Drug Safety eClinical Technologies Remedies for Medidata: Integrating Safer Drugs >> Infrastructure First Base for Clinical Good Days and Bad in 2011 >> Trials >> c CONTENTS | January 2012 FEATURE 46 Drug Safety Remedies for Safer Drugs >> UP FRONT 8 Drug Discovery H3 Biomedicine: Health, Hope and Heaps of Japanese Funding >> 12 Data Visualization Picture This: Molecular Maya Puts Life in Life Science Animations >> 18 Genome Informatics Assembly Required: Assemblathon I Tackles Complex Genomes >> 24 High Performance Computing Products, Deployments, and News from Supercomputing 11 >> 26 News Briefs >> 28 INSIDE THE BOX The Unstable Equilibrium of the Bioinformatics Org Chart >> 32 THE BUSH DOCTRINE The Pharmaceutical Safety Data Problem >> 2 JANUARY 2011 Visit Us at www.Bio-ITWorld.com Download a PDF of This Issue CLICK HERE! ARTICLES 36 eClinical Technologies Medidata: Integrating Infrastructure for Clinical Trials >> 42 Diagnostics A QuantuMDx Leap for Handheld DNA Sequencing >> 44 Informatics Tools Accelrys’ Cheminformatics Solution in the Cloud >> KNOWLEDGE CENTER 54 New Products >> 56 Free Trials and Downloads >> 58 Upcoming Events >> 60 Complimentary Resources >> IN EVERY ISSUE 6 FIRST BASE Good Days and Bad in 2011 >> 66 RUSSELL TRANSCRIPT What is (Quantitative) Systems Pharmacology? >> 64 On Deck | 64 Contact Us | 65 Company/Advertiser Index >> CONNECT WITH US! Visit Us at www.Bio-ITWorld.com JANUARY 2011 3 19th International Moscone North Convention Center February 19-23 San Francisco, CA February 19-20 NEW-Symposia Targeting Cancer Stem Cells get Cloud Computing DisplayMolecular of Difficult Targets Med Point-of-Care Diagnostics Next-Generation Pathology inspired Pharma-Bio Partnering Forums Attend.
[Show full text]
Massively Parallel Programming in Statistical Optimization & Simulation
Proceedings of the 2014 Winter Simulation Conference A. Tolk, S. Y. Diallo, I. O. Ryzhov, L. Yilmaz, S. Buckley, and J. A. Miller, eds. MASSIVELY PARALLEL PROGRAMMING IN STATISTICAL OPTIMIZATION & SIMULATION Russell Cheng Mathematical Sciences University of Southampton Building 54, Highfield Southampton, SO 17 1BJ, UNITED KINGDOM ABSTRACT General purpose graphics processing units (GPGPUs) suitable for general purpose programming have become sufficiently affordable in the last three years to be used in personal workstations. In this paper we assess the usefulness of such hardware in the statistical analysis of simulation input and output data. In particular we consider the fitting of complex parametric statistical metamodels to large data samples where optimization of a statistical function of the data is needed and investigate whether use of a GPGPU in such a problem would be worthwhile. We give an example, involving loss-given-default data obtained in a real credit risk study, where use of Nelder-Mead optimization can be efficiently implemented using parallel processing methods. Our results show that significant improvements in computational speed of well over an order of magnitude are possible. With increasing interest in "big data" samples the use of GPGPUs is therefore likely to become very important. 1 INTRODUCTION The use of parallel computing in discrete event simulation (DES) work has been well discussed in the literature. For a review see Fujimoto (1990). However it has only been in the last three years that hardware based on graphics processing units (GPUs) has become sufficiently widely available and affordable to make personal desktop parallel computing possible. A good general introduction to the field in general is given by Kirk and Hwu (2013).
[Show full text]
Interfacing with Opencl from Modern Fortran for Highly Parallel Workloads
Interfacing with OpenCL from Modern Fortran for Highly Parallel Workloads Laurence Kedward July 20 Copyright 2020 Laurence Kedward 1 Presentation Outline I. Background II. Focal – A pure Fortran interface library for OpenCL III. Results and Conclusions July 20 Copyright 2020 Laurence Kedward 2 I. Background Current and future trends in HPC Programming CPUs vs GPUs OpenCL overview July 20 Copyright 2020 Laurence Kedward 3 Maximising throughput: CPUs vs GPUs GPUs: maximise bandwidth • Highly parallel: 1000s of simple cores Latency: time- • Very high bandwidth memory cost per single work unit 퐁퐚퐧퐝퐰퐢퐝퐭퐡 Parallelism 퐓퐡퐫퐨퐮퐠퐡퐩퐮퐭 = 퐋퐚퐭퐞퐧퐜퐲 (amount of work Bandwidth: number of processed per unit work units processed time) concurrently Time CPUs: minimise latency • Large cache hierarchies (memory latency) • Instruction pipelining / branch prediction • Hyper-threading (hide latency) July 20 Copyright 2020 Laurence Kedward 4 Programming GPU architectures Problem type: SIMD • Running the same instructions on many thousands of different data • Little or no global synchronisation between threads Workload: arithmetic intensity • Maximise the amount of computation and minimise the amount of memory accessed Memory access: coalesced, structured, direct • Consecutive threads should access contiguous memory (no striding) • Hierarchical memory structure Physically distinct memory space: minimise host-device data transfer • PCIe bandwidth between GPU and CPU is very slow July 20 Copyright 2020 Laurence Kedward 5 Programming models Domain-specific Languages Libraries languages • CUDA (NVIDIA) • cuBLAS, cuFFT, etc. • PyFR • HIP (AMD) • clSPARSE • OP2 • OpenCL (Khronos) • Kokkos • OPS • SYCL (Khronos) • ArrayFire • OneAPI (Intel) • RAJA Directives • Loopy • OpenMP 4.0 • OpenACC July 20 Copyright 2020 Laurence Kedward 6 Programming models Domain-specific Languages Libraries languages • CUDA (NVIDIA) • cuBLAS, cuFFT, etc.
[Show full text]
Productive GPU Software a Working Microphone and Speakers
This conference uses integrated audio. To interact with the host, you need Productive GPU Software a working microphone and speakers. To speak, please click on the “Raise Hand” button in the Participants box. You can speak into your microphone after the host allows you to. If you cannot hear the host, or if your voice is not being transmitted, please let us know using the Chat window. Outline • Introduction to Jacket for MATLAB® • GFOR • Comparison with PCT™ alternative • Moving into the future • Case studies and code demos MATLAB® and Parallel Computing Toolbox™ (PCT) are trademarks of MathWorks® Easy GPU Acceleration of M code n = 20e6; % 20 million random samples X = grand(1,n,’gdouble’); Y = grand(1,n,’gdouble’); distance_to_origin = sqrt( X.*X + Y.*Y ); is_inside = (distance_to_origin <= 1); pi = 4 * sum(is_inside) / n; Matrix Types gdouble glogical double precision boolean guint# gint# unsigned integers integers gsingle single precision Matrix Types: ND Support vectors matrices volumes … ND Matrix Types: Easy Manipulation A(:,:,2) A(1,1) A(1,:) A(end,1) A(end,:) Easy GPU Acceleration of M code n = 20e6; % 20 million random samples X = grand(1,n); Y = grand(1,n); distance_to_origin = sqrt( X.*X + Y.*Y ); is_inside = (distance_to_origin <= 1); pi = 4 * sum(is_inside) / n; Easy GPU Acceleration of M code No GPU-specific stuff involved (no kernels, no threads, no blocks, just regular M code) “Very little recoding was needed to promote our Lattice Boltzmann Model code to run on the GPU.” –Dr. Kevin Tubbs, HPTi GFOR – Parallel FOR-loop
[Show full text]
A Practical Guide to Parallelization in Economics∗
A Practical Guide to Parallelization in Economics∗ Jesús Fernández-Villaverdey David Zarruk Valenciaz April 22, 2018 Abstract This guide provides a practical introduction to parallel computing in economics. After a brief introduction to the basic ideas of parallelization, we show how to paral- lelize a prototypical application in economics using, on CPUs, Julia, Matlab, R, Python, C++–OpenMP, Rcpp–OpenMP, and C++–MPI, and, on GPUs, CUDA and OpenACC. We provide code that the user can download and fork, present comparative results, and explain the strengths and weaknesses of each approach. We conclude with some additional remarks about alternative approaches. Key words: Computational Methods, Parallel Computing, Programming Languages. JEL classifications: C63, C68, E37. ∗The pdf file of the paper includes several embedded links of interest. The Github repository of the code used in the paper is: https://github.com/davidzarruk/Parallel_Computing. We thank many cohorts of students at the University of Pennsylvania who have endured earlier drafts of the slides behind this guide and pointed out ways to improve them. yUniversity of Pennsylvania. Email: [email protected] zUniversity of Pennsylvania. Email: [email protected] 1 Contents 1 Introduction4 2 Why Parallel?7 3 What Is Parallel Programming? 10 4 When Do We Parallelize? 12 4.1 An application where parallelization works: Value function iteration..... 13 4.2 An application where parallelization struggles: A random walk Metropolis- Hastings...................................... 15 4.3 The road ahead.................................. 18 5 A Life-Cycle Model 18 6 Different Computers 23 7 Parallelization Schemes on CPUs 25 7.1 Julia - Parallel for ............................... 26 7.2 Julia - MapReduce ................................
[Show full text]