ACCELERATING MATLAB SLOW LOOP EXECUTION WITH CUDA.PDF

Abstract

MATLAB is a high level language and computing environment widely used to write technical code. MATLAB is extensively used in a number of scientific fields such as digital image processing, digital signal processing etc. The significance of MATLAB programming is its rich library of built-in functions and ease of code writing. It offers a simple and easy way to write programming syntax, hence the learning curve is relatively small. However, MATLAB uses an interpreter which makes it slower, specially while executing loops. This slowness becomes a performance bottleneck in programs which use loops massively. In the field of digital image processing, for example, matrix operations usually involve many nested loop structures. On the other hand, MATLAB provides by the means of MEX files, interface to other programming languages like C and FORTRAN to be used in MATLAB. NVIDIA’s parallel computing architecture CUDA (Compute Unified Device Architecture) is a revolutionary parallel computing paradigm which utilizes GPUs for general purpose programming. CUDA uses massively parallel thread architecture and provides a programming interface in which threads are executed regardless of which CUDA enabled GPU they are running on. This makes a CUDA program easily scalable to CUDA enable GPUs with higher number of processing cores. A program in CUDA is written in either in C for CUDA (an extension of C language with CUDA) or in CUDA Driver API which is a low level programming paradigm. NVIDIA provides a MATLAB CUDA plug-in which is used to interface CUDA with MATLAB. By the means of this plug-in, it is possible to call functions written in CUDA into MATLAB like any other MATLAB function. The work carried out in this thesis is to investigate the benefits of CUDA architecture in MATLAB, in order to accelerate MATLAB’s slower processing. The main focus of our research is on MATLAB’s loops which are the main reason of MATLAB’s performance bottleneck. We used CUDA’s parallel processing architecture for simultaneous execution of each iteration of loops and hence accelerating the overall execution. We take an application example of JPEG compression and implement it on CUDA. We present a brief analysis of different parts of the algorithm in order to demonstrate which parts cause the performance bottleneck and how they can be executed in parallel on GPU. As a proof of concept, we provide different benchmarking results of CPU execution vs. GPU execution along with a brief discussion on the quality of x results. Our results show that CUDA can be successfully used in order to harness the slow execution