Compute Unified Device Architecture

vignette|upright=1.5|Schéma de principe de l'architecture CUDA CUDA (initialement l'acronyme de Compute Unified Device Architecture) est une technologie de GPGPU (General-Purpose Computing on Graphics Processing Units), c'est-à-dire utilisant un processeur graphique (GPU) pour exécuter des calculs généraux à la place du processeur central (CPU). En effet, ces processeurs comportent couramment de l'ordre d'un millier de circuits de calcul fonctionnant typiquement à , ce qui représente un potentiel très supérieur à un processeur central à , fût-il multicœurs et multi-threads. Cette utilisation requiert que le calcul à effectuer soit parallélisable. CUDA permet de programmer des GPU en C++. Elle est développée par Nvidia, initialement pour ses cartes graphiques GeForce 8 Series, et utilise un pilote unifié utilisant une technique de streaming (flux continu). Le premier kit de développement pour CUDA est publié le . Applications Est réalisable en CUDA tout ce qui demande du

GPU-accelerated Convex Multi-phase Image Segmentation

Xavier Bresson, Hadrien Copponnex, Arnaud Le Carvennec, Jean-Philippe Thiran, Subrahmanyam Venkata Ravi Mohana Sai Gorthi

Image segmentation is a key area of research in computer vision. Recent advances facilitated reformulation of the non-convex multi-phase segmentation problem as a convex optimization problem (see for example [2, 4, 9, 10, 13, 16]). Recently, [3] proposed a new convex relaxation approach for a class of vector-valued minimization problems, and this approach is directly applicable to the widely used classical Mumford-Shah segmentation model [11]. While the approach in [3] provides the much deserved convexification, it achieves this at the expense of an increased computational complexity due to the increased dimensionality of the reformulated problem; however, the algorithm proposed in [3] can indeed profit from a parallelized implementation. In this paper, we present a GPU-based implementation of the convex formulation for Mumford-Shah piecewise constant multi-phase image segmentation algorithm proposed in [3]. The main goal of this paper is to provide insights into the way the algorithm has been parallelized in order to obtain good speedup. We present multi-phase segmentation results both on synthetic and real images. The speedup of GPU based implementation is evaluated on three different GPUs. For sufficiently large images, the speedup achieved on GTX 285 GPU is around 40, compared to an optimized CPU implementation. The speedups obtained from GPU-based implementation are quite satisfactory. We also made our CUDA code available online.

2011

Massively parallel data processing for quantitative total flow imaging with optical coherence microscopy and tomography

Theo Lasser, Paul James Marchand, Marcin Antoni Sylwestrzak, Daniel Pawel Szlag

We present an application of massively parallel processing of quantitative flow measurements data acquired using spectral optical coherence microscopy (SOCM). The need for massive signal processing of these particular datasets has been a major hurdle for many applications based on SOCM. In view of this difficulty, we implemented and adapted quantitative total flow estimation algorithms on graphics processing units (GPU) and achieved a 150 fold reduction in processing time when compared to a former CPU implementation. As SOCM constitutes the microscopy counterpart to spectral optical coherence tomography (SOCT), the developed processing procedure can be applied to both imaging modalities. We present the developed DLL library integrated in MATLAB (with an example) and have included the source code for adaptations and future improvements. Program summary Program title: CudaOCMproc Catalogue identifier: AFBT_v1_0 Program summary URL: http://cpc.cs.qub.ac.uk/summaries/AFBT_v1_0.html Program obtainable from: CPC Program Library, Queen's University, Belfast, N. Ireland Licensing provisions: GNU GPLv3 No. of lines in distributed program, including test data, etc.: 913552 No. of bytes in distributed program, including test data, etc.: 270876249 Distribution format: tar.gz Programming language: CUDA/C, MATLAB. Computer: Intel x64 CPU, GPU supporting CUDA technology. Operating system: 64-bit Windows 7 Professional. Has the code been vectorized or parallelized?: Yes, CPU code has been vectorized in MATLAB, CUDA code has been parallelized. RAM: Dependent on users parameters, typically between several gigabytes and several tens of gigabytes Classification: 6.5, 18. Nature of problem: Speed up of data processing in optical coherence microscopy Solution method: Utilization of GPU for massively parallel data processing Additional comments: Compiled DLL library with source code and documentation, example of utilization (MATLAB script with raw data) Running time: 1,8 s for one B-scan (150 x faster in comparison to the CPU data processing time) (C) 2017 Published by Elsevier B.V.

Elsevier2017

Buffer overlow vulnerabilities in CUDA: a preliminary analysis

Andrea Miele

We present a preliminary study of buffer overflow vulnerabilities in CUDA software running on GPUs. We show how an attacker can overrun a buffer to corrupt sensitive data or steer the execution flow by overwriting function pointers, e.g., manipulating the virtual table of a C++ object. In view of a potential mass market diffusion of GPU accelerated software this may be a major concern.

Springer France2015

Massively parallel data processing for quantitative total flow imaging with optical coherence microscopy and tomography

Theo Lasser, Paul James Marchand, Marcin Antoni Sylwestrzak, Daniel Pawel Szlag

Elsevier2017