• KIPS Transactions on Computer and Communication Systems
• /
• v.8 no.5
• /
• pp.111-118
• /
• 2019

Recent graphic processing units (GPUs) provide high throughput by using powerful hardware resources. However, massive memory accesses cause GPU performance degradation due to cache inefficiency. Therefore, the performance of GPU can be improved by reducing thread parallelism when cache suffers memory contention. In this paper, we propose a dynamic warp scheduler which controls thread parallelism according to degree of cache contention. Usually, the greedy then oldest (GTO) policy for issuing warp shows lower parallelism than loose round robin (LRR) policy. Therefore, the proposed warp scheduler employs the LRR warp scheduling policy when Miss Status Holding Register(MSHR) utilization is low. On the other hand, the GTO policy is employed in order to reduce thread parallelism when MSHRs utilization is high. Our proposed technique shows better performance compared with LRR and GTO policy since it selects efficient scheduling policy dynamically. According to our experimental results, our proposed technique provides IPC improvement by 12.8% and 3.5% over LRR and GTO on average, respectively.

• The KIPS Transactions:PartB
• /
• v.19B no.4
• /
• pp.243-248
• /
• 2012

In the information retrieval systems like vector model implementation and document clustering, document similarity calculation takes a great part on the overall performance of the system. In this paper, GPU parallelism has been explored to enhance the processing speed of document similarity calculation in a CUDA framework. The proposed method increased the similarity calculation speed almost 15 times better compared to the typical CPU-based framework. It is 5.2 and 3.4 times better than the methods by using CUBLAS and Thrust, respectively.

• Journal of Digital Contents Society
• /
• v.9 no.3
• /
• pp.371-377
• /
• 2008

Recently the performance of GPU is increasing much faster compared to GPU and GPU is used for various application programs. In this paper, MPEG-2 Decoder is implemented based on a GPU programming language, CG. The proposed methodology is to perform block rendering with texture data according to video standard with very high parallelism by using the pipeline of GPU which is a stream processing structure. To reduce the data bandwidth between system memory and GPU, local memory is used for graphic card. According to the experiment, the proposed scheme shows performance improvement by more than 2 times compared to CPU based scheme.

• Journal of the Institute of Electronics Engineers of Korea SD
• /
• v.49 no.8
• /
• pp.1-9
• /
• 2012

Quantum-Inspired Evolutionary Algorithm(QEA) contains sufficient data-level parallelism to be naturally accelerated on GPUs. For an efficient reduction of execution time, however, careful task-mapping should be done to properly reflect the characteristics of CPU and GPU. Furthermore, when deciding which part of the application should run on GPU, we need to consider the data transfer between CPU and GPU memory spaces as well as the data-level parallelism. In addition, the usage of zero-copy host memory, proper choice of the execution configuration, and thread organization considering memory coalescing is important to further reduce the execution time. With all these techniques, we could run QEA 3.69 times faster on average in comparison with the multi-threading CPU for the case of 0-1 knapsack problem with 30,000 items.

• Journal of Korea Multimedia Society
• /
• v.19 no.8
• /
• pp.1297-1309
• /
• 2016

In this paper, an efficient connected component labeling (CCL) method was proposed. The proposed method is based on GPU parallelism. The CCL is very important in various applications where images are analysed. However, the label of each pixel is dependent on the connectivity of adjacent pixels so that it is not very easy to be parallelized. In this paper, a GPU-based parallel CCL techniques were proposed and applied to the analysis of radar signal. Since the radar signals contains complex and large data, the efficiency of the algorithm is crucial when realtime analysis is required. The experimental results show the proposed method is efficient enough to be successfully applied to this application.

• Journal of KIISE:Computing Practices and Letters
• /
• v.15 no.6
• /
• pp.445-449
• /
• 2009

In this paper, we propose a computationally efficient multi GPU accelerated image registration technique to correct the motion difference between the pre-contrast CT image and post-contrast CTA image. Our method consists of two steps: multi GPU based image registration and a cerebrovascular visualization. At first, it computes a similarity measure considering the parallelism between both GPUs as well as the parallelism inside GPU for performing the voxel-based registration. Then, it subtracts a CT image transformed by optimal transformation matrix from CTA image, and visualizes the subtracted volume using GPU based volume rendering technique. In this paper, we compare our proposed method with existing methods using 5 pairs of pre-contrast brain CT image and post-contrast brain CTA image in order to prove the superiority of our method in regard to visual quality and computational time. Experimental results show that our method well visualizes a brain vessel, so it well diagnose a vessel disease. Our multi GPU based approach is 11.6 times faster than CPU based approach and 1.4 times faster than single GPU based approach for total processing.

• Journal of the Korea Society of Computer and Information
• /
• v.27 no.2
• /
• pp.1-8
• /
• 2022

In this paper, we design and implement a GPU-based parallel algorithm to effectively solve the traveling salesman problem through an ant color system. The repetition process of generating hundreds or thousands of tours simultaneously in TSP utilizes GPU's task-level parallelism, and the update process of pheromone trails data actively exploits data parallelism by 32x32 thread blocks. In particular, through simultaneous memory access of multiple threads, the coalesced accesses on continuous memory addresses and concurrent accesses on shared memory are supported. This experiment used 127 to 1002 city data provided by TSPLIB, and compared the performance of sequential and parallel algorithms by using Intel Core i9-9900K CPU and Nvidia Titan RTX system. Performance improvement by GPU parallelization shows speedup of about 10.13 to 11.37 times.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2023.05a
• /
• pp.11-12
• /
• 2023

GPGPU 환경에서의 ML 모델이 다양한 분야에 지속적으로 활용되면서, 이미지 분할(image segmentation) 연구가 활발하다. multi-GPU 환경에서 성능 최적화를 위하여 병렬화 기법들이 활용되고 있다. 본 연구에서는 multi-GPU 환경에서 U-Net 모델의 전체 수행 시간을 단축하기 위해 convolution 연산을 최적화하는 기법을 적용하는 실험을 진행하였고 shared memory, data parallelism 를 적용하여 82% 성능 향상을 보여주었다.

• Annual Conference on Human and Language Technology
• /
• 2012.10a
• /
• pp.119-120
• /
• 2012

GPU를 이용한 병렬 알고리즘은 어떤 메모리를 사용하는지에 따라 시스템의 전체적인 성능이 달라진다. 본 논문은 GPU 환경에서 실행되는 CUDA 프레임워크에서 병렬처리를 이용하여 문서 분류 시스템의 속도를 향상시키고자 할 때 메모리 로딩 시간이 전체적인 시스템의 성능에 미치는 영항을 연구하였다. 기존의 CPU 환경에서 구현했을 때와 비교하여 어느 정도의 성능 향상이 있었는지 실험하였으며 이전 연구에서 고려하지 않았던 메모리를 읽는데 걸리는 시간을 고려하여 현실적인 실행 시간을 비교하였다. 실험 결과에 의하면 CPU 에서 구현했을 때의 연산 속도보다 GPU의 텍스쳐 메모리를 사용하여 구현하였을 때 문서분류 성능이 향상되는 효과가 있음을 알 수 있었다.

• Journal of the Korea Society of Computer and Information
• /
• v.24 no.4
• /
• pp.1-9
• /
• 2019

LRR(Loose Round Robin) warp scheduling policy for GPU architecture results in high warp-level parallelism and balanced loads across multiple warps. However, traditional LRR policy makes multiple warps execute long latency operations at the same time. In cases that no more warps to be issued under long latency, the throughput of GPUs may be degraded significantly. In this paper, we propose a new warp scheduling policy which utilizes latency hiding, leading to more utilized memory resources in high performance GPUs. The proposed warp scheduler prioritizes memory instruction based on GTO(Greedy Then Oldest) policy in order to provide reduced memory stalls. When no warps can execute memory instruction any more, the warp scheduler selects a warp for computation instruction by round robin manner. Furthermore, our proposed technique achieves high performance by using additional information about recently committed warps. According to our experimental results, our proposed technique improves GPU performance by 12.7% and 5.6% over LRR and GTO on average, respectively.