Implementation of Parallelization Techniques for PIM-Based JPEG Decoder in High-Resolution Image Processing

초록

JPEG is a digital image format widely used in various applications due to its ability to effectively reduce image file size. JPEG decoding involves restoring compressed image data for display or post-processing and plays a crucial role in fields such as real-time applications and deep learning, where large-scale image processing is required. However, JPEG decoding can become a performance bottleneck, prompting the development of various accelerator-based parallel processing techniques. Recently, research utilizing Processing-in-Memory (PIM) architectures has gained attention. Existing studies on PIM-based JPEG decoders have applied hardware threads and the AAN algorithm to improve performance, but limitations such as constrained memory capacity per DPU and lack of inter-DPU communication have restricted the resolution of image that can be processed. In this paper, we propose a modified JPEG decoding pipeline to overcome these limitations. By performing Huffman decoding in the host CPU, the need for inter-DPU communication is eliminated, enabling parallel processing without data dependencies between MCUs (Minimum Coded Units). Experiments conducted on the UPMEM PIM server using a high-resolution image dataset show that the proposed decoder can successfully decode high-resolution images that the existing PIM decoder could not process.

키워드