EXPLORING GPU MEMORY PERFORMANCE USING DIGITAL IMAGE PROCESSING ALGORITHMS
Journal Title: Indian Journal of Computer Science and Engineering - Year 2014, Vol 5, Issue 6
Abstract
Leveraging the incredible parallel computational power of graphics processing units (GPUs) is a proven method for accelerating general applications. Efficient utilization of the GPU remains one of the greatest challenges facing programmers. The performance of GPU applications is extremely reliant on memory performance, to the point that it can be considered a critical bottleneck. This is further amplified when working with large amounts of data, which is common. In this paper, we explore several well-known data transfer and memory access methods. Our aim is to find out how they affect the performance of different applications. To do so, we first examine and specify the different techniques; then, we apply these techniques to a variety of digital image processing applications, which serve as the case study. The NVIDIA CUDA parallel programming framework serves as the foundation for our research. Our experimental results highlight the merits of each optimization method. We then use these results to categorize the benchmarks according to their behavior. We demonstrate significantly superior performance including speedups of up to 24x compared to naïve implementations and up to 157x compared to serial implementations.
Authors and Affiliations
Puya Memarzia , Farshad Khunjush
Keywords
Related Articles
- EP ID EP105711
- DOI -
- Views 178
- Downloads 0
DETERMINING THE NUMBER OF CLUSTERS FOR A K-MEANS CLUSTERING ALGORTIHM
Clustering is a process used to divide data into a number of groups. All data points have some mathematical parameter according to which grouping can be done. For instance, if we have a number of points on a twodimension...
TEST SUITE GENERATION PROCESS FOR AGENT TESTING
Software agents are a promising technology for today's complex, distributed systems. Methodologies and techniques that address testing and reliability of multi agent systems are increasingly demanded, in particular to su...
Signed Numbers Conversions
Signed integers are normally represented using 2’s complement representation. Addition and subtraction of signed numbers is done in the same manner as for unsigned numbers. However carry (or borrow) is simple ignored. Un...
Efficient Resource Scheduling in Data Centers using MRIS
Scheduling the resources is decisive in shared data centers. Today scheduling algorithms focus only on onedimensional resource models, infact the multiple resources (e.g. Memory, Storage, CPU and Network bandwidth) can b...
AN IMPROVED DOMAIN CLASSIFICATION SCHEME BASED ON LOCAL FRACTAL DIMENSION
In fractal image compression, most of the time during encoding is spent for finding the best matching pair of range-domain blocks. Different techniques have been analyzed for decreasing the number of operations required...