CUDA Programming, GPU Computing

Graphics Processing Units (GPUs) are specialized coprocessors that were initially conceived for the purpose of accelerating vector operations, such as graphics rendering. Writing and configuring efficient algorithms for GPU devices is still a hard problem. The Algorithm Selection Problem consists of finding a combination of algorithms, or a configuration of an algorithm, that optimizes the solution of a given problem instance or set of instances. An auto-tuner is a program solves the Algorithm Selection Problem automatically. In this paper we implement an autotuner for the compilation flags of GPU algorithms, using the OpenTuner framework. The autotuner produces a set of compilation flags that aims to optimize the time to solve a given problem for a specific GPU device. We analyse the performance gains of tuning the compilation flags for heterogeneous GPU algorithms across three different GPU devices. We show that it is possible to gain performance by automatically and empirically selecting a set of compilation flags for the same GPU algorithm in different devices. In one of the experimental settings we were able to achieve a 30% speedup in comparison with the compiler high-level optimization options.

The computational epidemiology is the development and use of computational models that aims to understand the proliferation of diseases of the dynamic point of view. The computational models are capable to simulate the behavior of an epidemic and its effects on the population of a region and develop strategies of control and prevention of diseases. The computational epidemiology proposes models that classify individuals as susceptible, infected and recovered (SIR) and based on individual (IBM). The SIR model considers the homogeneous distribution of individuals in space and time and it is not capable to tell the persistence or eradication of infectious diseases. The IBM reproduces the premises of the SIR model to the individual analyse. The model computational cost to simulate the behavior of an epidemic grows up with the size of the population considered. This work implements a IBM using parallel programming through the use of GPU’s (Graphics Processor Unit) compatible with CUDA (Compute Unified Device Architecture). The goal is to gain performance compared to the traditional implementation. The results show that in some cases the reduction of computational time is around 170% and the growth rate with the increase of the population size is, in some cases, two times lower in a parallel implementation when compared to a sequencial implementation.

CUDA (Compute Unified Device Architecture) is a parallel computing platform developed by Nvidia which provides the ability of using GPUs to run computationally intensive programs. This presentation provides a brief overview of CUDA, parallelism in the GPU and the syntax needed to know to run the program on GPU.

Volume rendering is an important area of study in computer graphics, due to its application in areas such as medicine, physic simulations, oil and gas industries, and others. The main used method nowadays for volume rendering is ray casting. Nevertheless, there are a variety of parallel APIs that can be used to implement it. Thus, it is important to evaluate the performance of ray casting in different parallel APIs to help programmers in selecting one of them. In this paper, we present a performance comparison using OpenGL ® with fragment shader, OpenGL ® with compute shader, OpenCL, and CUDA.

This thesis focuses on the development, implementation and optimization of pattern-matching algorithms in two different, yet closely-related research fields: malicious code detection in intrusion detection systems and digital forensics (with a special focus on the data recovery process and the metadata collection stages it involves). The thesis introduces the motivational backgrounds for the development of the work, later on presents the related work and then continues with the main achievements obtained, while in the end a few conclusions and future research directions are discussed.

The main four chapters in this thesis show the main contributions of our work and address the following topics: we present an efficient storage mechanism for hybrid CPU/GPU-based systems, and compare it with other known approaches to date. We then propose an innovative, highly parallel approach to the fast construction of very large Aho-Corasick and Commentz-Walter pattern matching automata on hybrid CPU/GPU-based systems, and compare it to existing sequential approaches. Later on we propose a new heuristics for profiling malicious behavior based on system-call analysis, using the Aho-Corasick algorithm, and also discuss a new hybrid compression mechanism for this automata based on dynamic programming, that reduces the storage space required for it. Finally, we propose an efficient new method for collecting metadata and helping the human operator or automated tools used in the data recovery process as part of the computer forensic investigations.

The research and models obtained in this thesis extend the existing literature in the field of intrusion detection systems (malicious code detection in particular), by presenting: an innovative heuristics for behavioral analysis of code in executable files through system-call interception, a novel and highly efficient approach to efficiently storing pattern-matching automata in hybrid CPU/GPU-based systems, which serves as the base for an innovative model for the fast, GPU-accelerated construction of such very large automata (for both the Aho-Corasick and Commentz-Walter algorithms) and a new hybrid compression technique applied to the Aho-Corasick automata using a dynamic programming approach, that reduces storage space significantly.

The Graphics Processing Unit (GPU) has become an integral part of mainstream computing. The advancement and evaluation of GPU has lead to a significant performance improvement of many algorithms used in day today life. Being powerful Graphics engine, GPU’s are also powerful programmable processors featuring wide range of benefits over memory and arithmetic calculation. Along with GPU’s, advancement of Sensor Technology has led to many innovative Applications in recent years. The introduction of Microsoft Kinect has opened a plethora of application in Augmented Reality (AR) domain. Ornamental Dressing Room is an AR application meant for Jewellery. Aim of the application is to develop an Augmented Reality (AR) system that gives a visualization of ornaments worn on a human body. The User who intended to purchase ornaments, will face the Kinect Sensor, the application does the rest by identifying the control point at the neck and augments the selected jewelry over the neck. This project is aimed at comparing the results of applying parallel programming concepts for algorithms used in Ornament Dressing Room with GPU over CPU.

A flow diagram is a graphical presentation of an energy or chemical system with its components and their interconnections through mass and energy streams. An automatic drawing algorithm of flow diagrams has been developed and presented in this article. It heuristically imitates all the procedures carried out by a designer, starting with his conceptual understanding of the system’s topological structure and finishing with graphically representing the system on the paper or screen. The topological structure of the system is given as input in the form of digraph. As a first step, it is transformed to a planar digraph by introducing new vertices representing crossings between the various streams. The near-optimum (smallest) number of crossings is determined heuristically. Then the flow diagram is drawn on the screen using special mechanical engineering symbols for the components of the system. Horizontal and vertical lines represent streams of mass or energy, identified by different colors. Unnecessary expansion of the drawing area is avoided by the application of linear and integer-linear programming algorithms.

This paper presents a parallel implementation of the hybrid BiCGStab(2) (bi-conjugate gradient stabilized) iterative method in a GPU (graphics processing unit) for solution of large and sparse linear systems. This implementation uses the CUDA-Matlab integration, in which the method operations are performed in a GPU core using Matlab built-in functions. The goal is to show that the exploitation of parallelism by using this new technology can provide a significant computational performance. For the validation of the work, we compared the proposed implementation with a BiCGStab(2) sequential and parallelized implementation in the C and CUDA-C languages. The results showed that the proposed implementation is more efficient and can be viable for simulations being carried out with quality and in a timely manner. The gains in computational efficiency were 76x and 6x compared to the implementation in C and CUDA-C, respectively.

Fluid simulation has recently become possible to do in real-time utilizing modern programmable GPUs. In this paper, a comparison is made between more traditional CFD methods which solve the Navier-Stokes PDEs and the alternative discrete Lattice Boltzmann methods in this context. A practical imple- mentation was created by combining the two in an application that simulates snowfall affected by wind. The SOR-based solver for PDEs was found to be slightly faster on the grid resolutions plausible for a real-time simulation, significantly faster on large problem sizes (up to 131% faster), but slower on very small prob- lem sizes (LBM 31% faster).

This short paper present a collection of GPU lightweight decompression algorithms implementations within a FOSS library, Giddy – the first to be published to offer such function-ality. As the use of compression is important in ameliorating PCIe data transfer bottlenecks, we believe this library and its constituent implementations can serve as useful building blocks in GPU-accelerated DBMSes — as well as other data-intensive systems. The paper also includes an initial exploration of GPU-oriented patched compression schemes. Patching makes compression ratio robust against outliers, and is important with real-life data, which (in contrast to many synthetic benchmark datasets) exhibits non-uniform data distributions and noise. An experimental evaluation of both the unpatched and the patched schemes in Giddy is included.

Volume rendering is an important area of study in computer graphics, due to its application in areas such as medicine, physic simulations, oil and gas industries, and others. The main used method nowadays for volume rendering is ray casting. Nevertheless, there are a variety of parallel APIs that can be used to implement it. Thus, it is important to evaluate the performance of ray casting in different parallel APIs to help programmers in selecting one of them. In this paper, we present a performance comparison using OpenGL ® with fragment shader, OpenGL ® with compute shader, OpenCL, and CUDA.

Umumnya, Transformasi Fourier Diskrit digunakan untuk mengolah citra dengan mengubah sinyal yang berdomain waktu / spasial ke bentuk sinyal berdomain frekuensi, atau sebaliknya, mengubah sinyal berdomain frekuensi ke sinyal berdomain waktu / spasial. Salah satu contoh implememntasinya ialah memfilter data/sinyal untuk memisahkan sinyal yang yang telah diubah ke dalam domain frekuensi sehingga diperoleh sinyal diinginkan berdasarkan nilai frekuensi yang diinginkan.  Persamaan ini terdiri dari Transformasi Fourier Diskrit Maju (Forward Discrete Fourier Transform) dan Transformasi Fourier Diskrit Balik (Inverse Dscrete Fourier Transform).  Persamaan ini jika dibandingkan pengeksekusian pada CPU, pengeksekusian persamaan DFT (Discrete Fourier Transform) pada GPU (Graphical Processing Unit) menunjukkan pemercepatan yang cukup signifikan. Dengan ukuran sinyal maksimal hingga 65536 sinyal, diperoleh pemercepatan hingga 1,5 kali dibanding dengan persamaan DFT yang dieksekusi pada CPU.

This work has the goal to study how an efficient deep packet inspection (DPI) algorithm may be implemented using the graphical processing unit (GPU) CUDA (Computer Unified Device Architecture) enabled boards existing in personal computers, and to analyze implementation efficiency. The following tasks have been analyzed: the parallelization of the pattern matching algorithm and the optimization of C code written for Nvidia compiler to obtain the best performance. The conclusion shows that CUDA technology represents a very attractive solution to implement DPI algorithms without the typically memory and complexity constraints.

Usage of multiple unmanned aerial vehicles (UAV) in a certain mission makes flight route planning more complicated and slower. In order to obtain better performance, in the literature, most of the researchers propose using evolutionary algorithms and artificial intelligence approaches based on heuristics as optimization techniques. In addition to this, parallel programming approaches increase the computation performance. Therefore, this study focuses to discuss and solve the route planning problem for multi-UAV systems by using optimization techniques based on an evolutionary algorithm: simulated annealing. The travel cost and execution time are downsized in this work by optimization on algorithm and code. We implemented CPU based parallel solution to compare results with the GPU-accelerated one. The efficiency and the effectiveness of our parallelized and optimized solution is demonstrated through simulations under different scenarios. The results show that our optimized GPU based parallel solution for route planning problem is up to 1.6 times faster than serial and parallel CPU solutions. Moreover, our optimized GPU solution is better on cost than other solutions. It is shown that our GPU based approach is the fastest one and increases performance thanks to the massive parallelization capabilities of GPUs. Index Terms-GPU programming, parallel programming, route planning, simulated annealing.

The last years have seen the rise of using GPU’s not only for graph- ics computing, but also for a more global use in parallel computing. This completely new field comes with dedicated GPU’s (GPGPU, Global Purpose GPU) that are not specialized anymore, but can be used to compute op- erations normally conducted on a CPU. Along with hardwares comes new programming technologies that enable programmers to use the GPU in their programs. This field is mainly divided between CUDA, which comes from NVIDIA, and OpenCL, which is an open standard. In this paper, we will be looking at technologies that enable programmers to perform GPU operations from Java code so as to evaluate them. The first chapter (Introduction) will select the candidates and set the test battery. The second one (Analysis) will evaluate the candidates according to our test battery. The third chapter (Results) will compare the results obtained between the candidates and the last chapter (Conclusion) will finally explain which solution is better and in which cases.

The rapid evolution of CUDA GPU architecture and the new heterogenous platforms that break the hegemony of x86 offer opportunities for performance optimizations, but also pose challenges for scalable heterogeneous parallelization of the GROMACS molecular simulation package. This session will present our latest efforts to harness recent CUDA architectures to improve algorithmic efficiency and performance of our molecular dynamics kernels. We will also discuss load balancing and latency-hiding challenges emphasized by the expansions of GPU-accelerated platforms with CPUs a ranging from a power-optimized ARM architectures to extreme-performance highly multi-threaded Power and Xeon CPUs. Come to learn about our experiences in developing portable heterogeneous high performance code!

Рассмотрен мониторинг широкого частотного диапазона с использованием ДПФ-модулированных банков фильтров. Описаны равнополосные и неравнополосные реализации банков фильтров, включая прямую реализа-цию с полной модуляцией, критически децимированную полифазную и передискретизированную полифазную реализации. Исследована точность восстановления сигнала каждой реализацией при различных отношениях “сигнал/шум” с использованием программных средств MATLAB. Рассмотрена аппаратная реализация ДПФ-модулированных банков фильтров на базе программируемых логических интегральных схем (ПЛИС) и графиче-ских вычислителей, использующих технологию CUDA.
Wideband monitoring with the use of DFT-modulated filter banks is discussed. Filter banks with equal and unequal bandwidths are described including direct forms with full modulation, critically decimated polyphase forms, and resampling polyphase forms. Experimental research of signal reconstruction accuracy for different signal-to-noise ratios in MATLAB is described. Hardware realization of DFT-modulated filter banks with the use the programmable logic integrated circuits and graphical computers based on CUDA is considered.

En este artículo se plantea una nueva paralelización del algoritmo de enjambre de partículas (PSO) haciendo uso de múltiples GPUs. Se plantean dos implementaciones basadas en la arquitectura CUDA de NVIDIA y la conexión P2P a través del puerto PCIe entre las GPUs.
Se evalúa la aceleración sobre dos GPUs comparando los resultados con la implementación secuencial y con una única GPU. Los resultados demuestran que usando el esquema clásico de PSO sobre un sistema multi-GPU la latencia de comunicación es alta, por tanto se usa una distribución en islas de procesamiento con el cual se logra una aceleración de hasta 32x respecto a la implementación secuencial.

This paper presents a computational performance comparison between some iterative methods used for linear systems solution. The goal is to show that the use of parallel processing provided by a Graphics Processing Unit (GPU) may be more feasible, for making possible the fast solution of linear equations systems in order that complex and sparse problems can be solved in a short time. To validate the paper a GPU through the NVIDIA's Compute Unified Device Architecture (CUDA) was employed and the computational performance was compared with Jacobi, Gauss-Seidel, BiCGStab iterative methods and BiCGStab(2) parallelized in the solution of linear systems of varying sizes. There was a significant acceleration in tests with the parallelized code, which increases considerably as much as systems increase. The results showed that the application of parallel processing in a robust and efficient method, as BiCGStab(2), it is often necessary for the simulations be performed with quality and in not prohibitive time.

The availability of Internet, line-of-sight and satellite identification and surveillance information as well as low-power, low-cost embedded systems-on-a-chip and a wide range of visible to long-wave infrared cameras prompted Embry Riddle Aeronautical University to collaborate with the University of Alaska Arctic Domain Awareness Center (ADAC) in summer 2016 to prototype a camera system we call the SDMSI (Software-Defined Multi-spectral Imager). The concept for the camera system from the start has been to build a sensor node that is drop-in-place for simple roof, marine, pole-mount, or buoy-mounts. After several years of component testing, the integrated SDMSI is now being tested, first on a roof-mount at Embry Riddle Prescott. The roof-mount testing demonstrates simple installation for the high spatial, temporal and spectral resolution SDMSI. The goal is to define and develop software and systems technology to complement satellite remote sensing and human monitoring of key resources such as drones, aircraft and marine vessels in and around airports, roadways, marine ports and other critical infrastructure. The SDMSI was installed at Embry Riddle Prescott in fall 2016 and continuous recording of long-wave infrared and visible images have been assessed manually and compared to salient object detection to automatically record only frames containing objects of interest (e.g. aircraft and drones). It is imagined that ultimately users of the SDMSI can pair with it via wireless to browse salient images. Further, both ADS-B (Automatic Dependent Surveillance-Broadcast) and S-AIS (Satellite Automatic Identification System) data are envisioned to be used by the SDMSI to form expectations for observing in future tests. This paper presents the preliminary results of several experiments and compares human review with smart image processing in terms of the receiver-operator characteristic. The system design and software are open architecture, such that other researchers are encouraged to construct and participate in sharing results and networking identical or improved versions of the SDMSI for safety, security and drop-in-place scientific image sensor networking. Nomenclature ADS-B = Automatic Dependent Surveillance – Broadcast, aviation identification and tracking AIFC = Arctic Information Fusion Concept, an ADAC sensor network prototype CBONS = Community Based Observing Network System, or human field monitoring CUDA = Compute Unified Device Architecture, GP-GPU acceleration DMM = Digital Multi-Meter, used for current monitoring and power use analysis EO/IR = Electro-Optical / Infrared instrumentation GPGPU = General Purpose Graphics Processing Unit LWIR = Long Wave Infrared, typically 8-15 micron wavelength electromagnetic radiation MWIR = Medium Wave Infrared, typically 3-8 micron wavelength electromagnetic radiation NAS = National Airspace NIR = Near Infrared, typically 0.75-1.4 micron wavelength electromagnetic radiation OpenCV = Open Computer Vision, an open source library in C/C++ Panchromatic = Visible and part of VNIR electromagnetic radiation in 0.45-0.8 micron range PCIe = Peripheral Component Interconnect Express, a device interface bus ROC = Receiver operator Characteristic, compares true positive and false positive rates S-AIS = Satellite Automatic Identification System – automatic marine tracking service SDMSI = Software Defined Multi-Spectral Imager SOD = Salient Object Detector SWIR = Short Wave Infrared, typically 1.4-3 micron wavelength electromagnetic radiation TAP = Trans Alaska Pipeline UAS = Unoccupied Aerial System USB3 = Universal Serial Bus, Revision 3, operating at 5 gigabits per second (625MB/sec) USCG = US Coast Guard VNIR = Visible and Near Infrared, typically 04.4-1 micron wavelength range Introduction The purpose of the research presented in this paper is to evaluate the hypothesis that pole-mount cameras on buoys, buildings or towers, and marine vessels can improve situational awareness for the agencies and organizations that manage campuses, ports, airports and other critical infrastructure where drone, aircraft and marine vessels cooperate compared to use of satellite remote sensing and human monitoring. The assertion is that a multi-spectral imaging system defined by software providing concurrent visible and infrared image collection and processing can also be defined and improved through software upgrades over time to perform better than security camera continuous monitoring or occasional satellite imaging. Finally, that the result will be better spatial, temporal, and spectral resolution observing of key areas of interest in regions that are hard to monitor such as Alaska and the Arctic compared to current methods employed. This hypothesis has been initially tested at Embry Riddle by monitoring shared airspace traffic including drones, aircraft and wildlife to test whether the concept of a smart SDMSI might also have value for aerial surveys and surveillance as well as marine environments. The SDMSI system design that has been prototyped and built and tested in Arizona is shown in Figure 1 below. The camera system includes a Tegra K1 SoC (4 processor

This paper presents a comparison between two parallel architectures: Compute Unified Device Architecture (CUDA) and Open Computing Language (OpenCL). Some works in the literature have presented a computational performance comparison of the two architectures. However, there is not some complete and recent paper that highlights clearly which architecture can be considered the most efficient. Thus, the goal is to make a comparison only in level of hardware, software, technological trends and ease of use, highlighting one that may present the best performance in general. To this end, we describe the main works that have used at least one of the architectures. It was observed that the choice of OpenCL may seem more obvious for being a heterogeneous system. Nevertheless, it was concluded that CUDA, although it can be used only in graphics cards from NVIDIA, has been a reference and more used recently.

Overview CUDA stands for the "Compute Unified Device Architecture", which is a free software platform provided by NVidia. It enables users to control GPUs by writing programs akin to C++. All CUDA software can be downloaded from CUDA Zone.CUDA is very similar in nature to cores in a CPU. You might have a dual or quad-core CPU, and that is the closest analog to CUDA cores that most people will have any experience with. However, CUDA cores tend to be more specialized for stream processing as opposed to generalized like a CPU, so there's less logic to duplicate making it easier to fit more CUDA cores onto a GPU. That said, CUDA's primary use seems to be for allowing the GPU to be used for more general purpose tasks, like encoding videos, or even hardware acceleration of video decoding. For gaming, I would go more for things like memory bus bandwidth, as well as memory and GPU clock speeds. If you can find a card that also has more CUDA cores compared to a very similar card, then even better, but I'd put CUDA cores a ways down on the list for gaming. Even just considering GPUs, while NVidia is the most popular, it's not the only game in town; there are also AMD's (ATI's). They use OpenCL, which the Kronos group labels. It is a different API for GPU-like (stream, SIMD) processing. OpenCL is a standard, and you can run OpenCL-written code on NVidia gear, but it won't run as fast as CUDA code. Using CUDA allows the programmer to take advantage of the massive parallel computing power of an NVidia graphics card in order to do general purpose computation. Before continuing, it's worth talking about this for a little bit longer. Multicore CPU and GPU CPUs like Intel Core 2 Duo and AMD Opteron are good at doing one or two tasks at a time, and doing those tasks very quickly. Graphics cards, on the other hand, are good at doing a massive number tasks at the same time, and doing those tasks relatively quickly. To put this into perspective, suppose you have a 20 inch monitor with a standard resolution of 1,920 x 1200. An NVidia graphics card has the computational ability to calculate the color of 2,304,000 different pixels, many times a second. In order to accomplish this feat, graphics cards use dozens, even hundreds of ALUs. Fortunately, NVidia's ALUs are fully programmable, which enables us to harness an unprecedented amount of computational power into the programs that we write. As stated previously, CUDA lets the programmer take advantage of the hundreds of ALUs inside a graphics processor, which is much more powerful than the handful of ALUs available in any CPU. However, this does put a limit on the types of applications that are well suited to CUDA.

This paper proposes a parallel regression formulation to reduce the computational time of variable selection algorithms. The proposed strategy can be used for several forward algorithms in order to select uncorrelated variables that contribute for a better predictive capability of the model. Our demonstration of the proposed method include the use of Successive Projections Algorithm (SPA), which is an iterative forward technique that minimizes multicollinearity. SPA is traditionally used for variable selection in the context of multivariate calibration. Nevertheless, due to the need of calculating an inverse matrix for each insertion of a new variable in the model calibration, the computational performance of the algorithm may become impractical as the matrix size increases. Based on such limitation, this paper proposes a new strategy called Parallel Regressions (PR). PR strategy was implemented in the SPA to avoid the matrix inverse calculation of original SPA in order to increase the computational performance of the algorithm. It uses a parallel computing platform called Compute Unified Device Architecture (CUDA) in order to exploit a Graphics Processing Unit, and was called SPA-PR-CUDA. For this purpose, we used a case study involving a large data set of spectral variables. The results obtained with SPA-PR-CUDA presented 37x times better performance compared to a traditional SPA implementation. Additionally, when compared to traditional algorithms we demonstrated that SPA-PR-CUDA may be a more viable choice for obtaining a model with a reduced prediction error value.