NVIDIA Compute Unified Device Architecture Research Articles

BarraCUDA - a fast short read sequence aligner using graphics processing units

BackgroundWith the maturation of next-generation DNA sequencing (NGS) technologies, the throughput of DNA sequencing reads has soared to over 600 gigabases from a single instrument run. General purpose computing on graphics processing units (GPGPU), extracts the computing power from hundreds of parallel stream processors within graphics processing cores and provides a cost-effective and energy efficient alternative to traditional high-performance computing (HPC) clusters. In this article, we describe the implementation of BarraCUDA, a GPGPU sequence alignment software that is based on BWA, to accelerate the alignment of sequencing reads generated by these instruments to a reference DNA sequence.FindingsUsing the NVIDIA Compute Unified Device Architecture (CUDA) software development environment, we ported the most computational-intensive alignment component of BWA to GPU to take advantage of the massive parallelism. As a result, BarraCUDA offers a magnitude of performance boost in alignment throughput when compared to a CPU core while delivering the same level of alignment fidelity. The software is also capable of supporting multiple CUDA devices in parallel to further accelerate the alignment throughput.ConclusionsBarraCUDA is designed to take advantage of the parallelism of GPU to accelerate the alignment of millions of sequencing reads generated by NGS instruments. By doing this, we could, at least in part streamline the current bioinformatics pipeline such that the wider scientific community could benefit from the sequencing technology.BarraCUDA is currently available from http://seqbarracuda.sf.net

Porting and optimizing MAGFLOW on CUDA

The MAGFLOW lava simulation model is a cellular automaton developed by the Sezione di Catania of the Istituto Nazionale di Geofisica e Vulcanologia (INGV) and it represents the peak of the evolution of cell-based models for lava-flow simulation. The accuracy and adherence to reality achieved by the physics-based cell evolution of MAGFLOW comes at the cost of significant computational times for long-running simulations. The present study describes the efforts and results obtained by porting the original serial code to the parallel computational platforms offered by modern video cards, and in particular to the NVIDIA Compute Unified Device Architecture (CUDA). A number of optimization strategies that have been used to achieve optimal performance on a graphic processing units (GPU) are also discussed. The actual benefits of running on the GPU rather than the central processing unit depends on the extent and duration of the simulated event; for large, long-running simulations, the GPU can be 70-to-80-times faster, while for short-lived eruptions with a small extents the speed improvements obtained are 40-to-50 times.

Guest Editor's Introduction: Special Section on Challenges and Solutions in Multicore and Many‐Core Computing

Guest Editor's Introduction: Special Section on Challenges and Solutions in Multicore and Many‐Core Computing

Real-Time Artifact-Free Image Upscaling

The problem of creating artifact-free upscaled images appearing sharp and natural to the human observer is probably more interesting and less trivial than it may appear. The solution to the problem, often referred to also as "single-image super-resolution," is related both to the statistical relationship between low-resolution and high-resolution image sampling and to the human perception of image quality. In many practical applications, simple linear or cubic interpolation algorithms are applied for this task, but the results obtained are not really satisfactory, being affected by relevant artifacts like blurring and jaggies. Several methods have been proposed to obtain better results, involving simple heuristics, edge modeling, or statistical learning. The most powerful ones, however, present a high computational complexity and are not suitable for real-time applications, while fast methods, even if edge adaptive, are not able to provide artifacts-free images. In this paper, we describe a new upscaling method (iterative curvature-based interpolation) based on a two-step grid filling and an iterative correction of the interpolated pixels obtained by minimizing an objective function depending on the second-order directional derivatives of the image intensity. We show that the constraints used to derive the function are related with those applied in another well-known interpolation method, providing good results but computationally heavy (i.e., new edge-directed interpolation (NEDI). The high quality of the images enlarged with the new method is demonstrated with objective and subjective tests, while the computation time is reduced of one to two orders of magnitude with respect to NEDI so that we were able, using a graphics processing unit implementation based on the nVidia Compute Unified Device Architecture technology, to obtain real-time performances.

Understanding the Efficiency of kD-tree Ray-Traversal Techniques over a GPGPU Architecture

Current GPU computational power enables the execution of complex and parallel algorithms, such as ray tracing techniques supported by kD-trees for 3D scene rendering in real time. This work describes in detail the study and implementation of eight different kD-tree traversal algorithms using the parallel framework NVIDIA Compute Unified Device Architecture, in order to point their pros and cons regarding performance, memory consumption, branch divergencies and scalability on multiple GPUs. In addition, two new algorithms are proposed by the authors based on this analysis, aiming to performance improvement. Both of them are capable of reaching speedup gains up to 3 × when compared to recent and optimized parallel traversal implementations. As a consequence, interactive frame rates are possible for scenes with 1,408 × 768 pixels of resolution and 3.6 million primitives.

ALICE HLT High Speed Tracking on GPU

The on-line event reconstruction in ALICE is performed by the High Level Trigger, which should process up to 2000 events per second in proton-proton collisions and up to 300 central events per second in heavy-ion collisions, corresponding to an input data stream of 30 GB/s. In order to fulfill the time requirements, a fast on-line tracker has been developed. The algorithm combines a Cellular Automaton method being used for a fast pattern recognition and the Kalman Filter method for fitting of found trajectories and for the final track selection. The tracker was adapted to run on Graphics Processing Units (GPU) using the NVIDIA Compute Unified Device Architecture (CUDA) framework. The implementation of the algorithm had to be adjusted at many points to allow for an efficient usage of the graphics cards. In particular, achieving a good overall workload for many processor cores, efficient transfer to and from the GPU, as well as optimized utilization of the different memories the GPU offers turned out to be critical. To cope with these problems a dynamic scheduler was introduced, which redistributes the workload among the processor cores. Additionally a pipeline was implemented so that the tracking on the GPU, the initialization and the output processed by the CPU, as well as the DMA transfer can overlap. The GPU tracking algorithm significantly outperforms the CPU version for large events while it entirely maintains its efficiency.

SU (2) lattice gauge theory simulations on Fermi GPUs

In this work we explore the performance of CUDA in quenched lattice SU (2) simulations. CUDA, NVIDIA Compute Unified Device Architecture, is a hardware and software architecture developed by NVIDIA for computing on the GPU. We present an analysis and performance comparison between the GPU and CPU in single and double precision. Analyses with multiple GPUs and two different architectures (G200 and Fermi architectures) are also presented. In order to obtain a high performance, the code must be optimized for the GPU architecture, i.e., an implementation that exploits the memory hierarchy of the CUDA programming model. We produce codes for the Monte Carlo generation of SU (2) lattice gauge configurations, for the mean plaquette, for the Polyakov Loop at finite T and for the Wilson loop. We also present results for the potential using many configurations (50,000) without smearing and almost 2000 configurations with APE smearing. With two Fermi GPUs we have achieved an excellent performance of 200× the speed over one CPU, in single precision, around 110 Gflops/s. We also find that, using the Fermi architecture, double precision computations for the static quark-antiquark potential are not much slower (less than 2× slower) than single precision computations.

GPU-accelerated indirect boundary element method for voxel model analyses with fast multipole method

An indirect boundary element method (BEM) that uses the fast multipole method (FMM) was accelerated using graphics processing units (GPUs) to reduce the time required to calculate a three-dimensional electrostatic field. The BEM is designed to handle cubic voxel models and is specialized to consider square voxel walls as boundary surface elements. The FMM handles the interactions among the surface charge elements and directly outputs surface integrals of the fields over each individual element. The CPU code was originally developed for field analysis in human voxel models derived from anatomical images. FMM processes are programmed using the NVIDIA Compute Unified Device Architecture (CUDA) with double-precision floating-point arithmetic on the basis of a shared pseudocode template. The electric field induced by DC-current application between two electrodes is calculated for two models with 499,629 (model 1) and 1,458,813 (model 2) surface elements. The calculation times were measured with a four-GPU configuration (two NVIDIA GTX295 cards) with four CPU cores (an Intel Core i7-975 processor). The times required by a linear system solver are 31 s and 186 s for models 1 and 2, respectively. The speed-up ratios of the FMM range from 5.9 to 8.2 for model 1 and from 5.0 to 5.6 for model 2. The calculation speed for element-interaction in this BEM analysis was comparable to that of particle-interaction using FMM on a GPU.

Real-time blood flow visualization using the graphics processing unit

Laser speckle imaging (LSI) is a technique in which coherent light incident on a surface produces a reflected speckle pattern that is related to the underlying movement of optical scatterers, such as red blood cells, indicating blood flow. Image-processing algorithms can be applied to produce speckle flow index (SFI) maps of relative blood flow. We present a novel algorithm that employs the NVIDIA Compute Unified Device Architecture (CUDA) platform to perform laser speckle image processing on the graphics processing unit. Software written in C was integrated with CUDA and integrated into a LabVIEW Virtual Instrument (VI) that is interfaced with a monochrome CCD camera able to acquire high-resolution raw speckle images at nearly 10 fps. With the CUDA code integrated into the LabVIEW VI, the processing and display of SFI images were performed also at ∼10 fps. We present three video examples depicting real-time flow imaging during a reactive hyperemia maneuver, with fluid flow through an in vitro phantom, and a demonstration of real-time LSI during laser surgery of a port wine stain birthmark.

Modeling and Rendering of Real-time Large-scale Granular Flow Scene on GPU

Real-time large-scale granular flow scene simulation has become a challenging task because massive amounts of geometric and graphical computation need to be rapidly processed. The paper presents a new simulation method that compute Discrete Element Method (DEM) entirely on GPU. To achieve this goal, Spatial subdivision algorithm is firstly proposed to partition space into a uniform grid, we also propose parallel radix sort algorithm and then sort the particles based on their hash values. With the above methods, we can effectively increase speed for searching neighboring particles and interacting with boundaries. To reach the requirement of real-time simulation of large-scale granular flow based on the Discrete Element Method (DEM), the implementation is made with the NVIDIA Compute Unified Device Architecture (CUDA). The experimental results show that the method proposed not only can simulate large-scale granular flow scene in real-time, but can simulate various kinds of granular flow behaviors in real-time by adjusting the very few experimental parameters.

CudaBayesreg: Parallel Implementation of a Bayesian Multilevel Model for fMRI Data Analysis

Graphic processing units (GPUs) are rapidly gaining maturity as powerful general parallel computing devices. A key feature in the development of modern GPUs has been the advancement of the programming model and programming tools. Compute Unified Device Architecture (CUDA) is a software platform for massively parallel high-performance computing on Nvidia many-core GPUs. In functional magnetic resonance imaging (fMRI), the volume of the data to be processed, and the type of statistical analysis to perform call for high-performance computing strategies. In this work, we present the main features of the R-CUDA package cudaBayesreg which implements in CUDA the core of a Bayesian multilevel model for the analysis of brain fMRI data. The statistical model implements a Gibbs sampler for multilevel/hierarchical linear models with a normal prior. The main contribution for the increased performance comes from the use of separate threads for fitting the linear regression model at each voxel in parallel. The R-CUDA implementation of the Bayesian model proposed here has been able to reduce significantly the run-time processing of Markov chain Monte Carlo (MCMC) simulations used in Bayesian fMRI data analyses. Presently, cudaBayesreg is only configured for Linux systems with Nvidia CUDA support.

GPU-accelerated 3-D model-based tracking

Model-based approaches to tracking the pose of a 3-D object in video are effective but computationally demanding. While statistical estimation techniques, such as the particle filter, are often employed to minimize the search space, real-time performance remains unachievable on current generation CPUs. Recent advances in graphics processing units (GPUs) have brought massively parallel computational power to the desktop environment and powerful developer tools, such as NVIDIA Compute Unified Device Architecture (CUDA), have provided programmers with a mechanism to exploit it. NVIDIA GPUs' single-instruction multiple-thread (SIMT) programming model is well-suited to many computer vision tasks, particularly model-based tracking, which requires several hundred 3-D model poses to be dynamically configured, rendered, and evaluated against each frame in the video sequence. Using 6 degree-of-freedom (DOF) rigid hand tracking as an example application, this work harnesses consumer-grade GPUs to achieve real-time, 3-D model-based, markerless object tracking in monocular video.

Scalability of Higher-Order Discontinuous Galerkin FEM Computations for Solving Electromagnetic Wave Propagation Problems on GPU Clusters

A highly parallel implementation of Maxwell's equations in the time domain using a cluster of Graphics Processing Units (GPUs) is presented. The higher-order Discontinuous Galerkin Finite Element Method (DG-FEM) is used for spatial discretization since its characteristics are matching the parallelization design aspects of the NVIDIA Compute Unified Device Architecture (CUDA) programming model. Asynchronous data transfer is introduced to minimize parallelization overhead and improve parallel efficiency. The implementation is benchmarked with help of a realistic 3-D geometry of an electromagnetic compatibility problem.

Real-time nonlinear finite element computations on GPU – Application to neurosurgical simulation

Application of biomechanical modeling techniques in the area of medical image analysis and surgical simulation implies two conflicting requirements: accurate results and high solution speeds. Accurate results can be obtained only by using appropriate models and solution algorithms. In our previous papers we have presented algorithms and solution methods for performing accurate nonlinear finite element analysis of brain shift (which includes mixed mesh, different non-linear material models, finite deformations and brain–skull contacts) in less than a minute on a personal computer for models having up to 50,000 degrees of freedom. In this paper we present an implementation of our algorithms on a graphics processing unit (GPU) using the new NVIDIA Compute Unified Device Architecture (CUDA) which leads to more than 20 times increase in the computation speed. This makes possible the use of meshes with more elements, which better represent the geometry, are easier to generate, and provide more accurate results.

Hybrid of genetic algorithm and local search to solve MAX-SAT problem using nVidia CUDA framework

General Purpose computing over Graphical Processing Units (GPGPUs) is a huge shift of paradigm in parallel computing that promises a dramatic increase in performance. But GPGPUs also bring an unprecedented level of complexity in algorithmic design and software development. In this paper we describe the challenges and design choices involved in parallelizing a hybrid of Genetic Algorithm (GA) and Local Search (LS) to solve MAXimum SATisfiability (MAX-SAT) problem on a state-of-the-art nVidia Tesla GPU using nVidia Compute Unified Device Architecture (CUDA). MAX-SAT is a problem of practical importance and is often solved by employing metaheuristics based search methods like GAs and hybrid of GA with LS. Almost all the parallel GAs (pGAs) designed in the last two decades were designed for either clusters or MPPs. Unfortunately, very little research is done on the implementation of such algorithms over commodity graphics hardware. GAs in their simple form are not suitable for implementation over the Single Instruction Multiple Thread (SIMT) architecture of a GPU, and the same is the case with conventional LS algorithms. In this paper we explore different genetic operators that can be used for an efficient implementation of GAs over nVidia GPUs. We also design and introduce new techniques/operators for an efficient implementation of GAs and LS over such architectures. We use nVidia Tesla C1060 to perform several numerical tests and performance measurements and show that in the best case we obtain a speedup of 25×. We also discuss the effects of different optimization techniques on the overall execution time.

Teraflop per second gravitational lensing ray-shooting using graphics processing units

Gravitational lensing calculation using a direct inverse ray-shooting approach is a computationally expensive way to determine magnification maps, caustic patterns, and light-curves (e.g. as a function of source profile and size). However, as an easily parallelisable calculation, gravitational ray-shooting can be accelerated using programmable graphics processing units (GPUs). We present our implementation of inverse ray-shooting for the NVIDIA G80 generation of graphics processors using the NVIDIA Compute Unified Device Architecture (CUDA) software development kit. We also extend our code to multiple GPU systems, including a 4-GPU NVIDIA S1070 Tesla unit. We achieve sustained processing performance of 182 Gflop/s on a single GPU, and 1.28 Tflop/s using the Tesla unit. We demonstrate that billion-lens microlensing simulations can be run on a single computer with a Tesla unit in timescales of order a day without the use of a hierarchical tree-code.