Canonical Huffman Decoder on Fine-grain Many-core Processor Arrays

Sarangi, Satyabrata; Baas, Bevan M.

doi:10.1145/3394885.3431424

Cited by 5 publications

(2 citation statements)

References 15 publications

(21 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Optimized Lossless Image Decoding. Many proposals attempt to improve the decoding throughput of the lossless image format, such as Huffman decoding [47,54,56], bzip2 [39], and LZSS [36] by utilizing GPUs or specialized hardware. However, there are trade-offs between decoding throughput and compression ratio.…”

Section: Related Workmentioning

confidence: 99%

L3: Accelerator-Friendly Lossless Image Format for High-Resolution, High-Throughput DNN Training

Bae¹,

Baek²,

Ham³

et al. 2022

Preprint

View full text Add to dashboard Cite

The training process of deep neural networks (DNNs) is usually pipelined with stages for data preparation on CPUs followed by gradient computation on accelerators like GPUs. In an ideal pipeline, the end-to-end training throughput is eventually limited by the throughput of the accelerator, not by that of data preparation. In the past, the DNN training pipeline achieved a near-optimal throughput by utilizing datasets encoded with a lightweight, lossy image format like JPEG. However, as high-resolution, losslessly-encoded datasets become more popular for applications requiring high accuracy, a performance problem arises in the data preparation stage due to low-throughput image decoding on the CPU. Thus, we propose L3, a custom lightweight, lossless image format for high-resolution, high-throughput DNN training. The decoding process of L3 is effectively parallelized on the accelerator, thus minimizing CPU intervention for data preparation during DNN training. L3 achieves a 9.29× higher data preparation throughput than PNG, the most popular lossless image format, for the Cityscapes dataset on NVIDIA A100 GPU, which leads to 1.71× higher end-to-end training throughput. Compared to JPEG and WebP, two popular lossy image formats, L3 provides up to 1.77× and 2.87× higher end-to-end training throughput for ImageNet, respectively, at equivalent metric performance.

show abstract

Section: Related Workmentioning

confidence: 99%

L3: Accelerator-Friendly Lossless Image Format for High-Resolution, High-Throughput DNN Training

Bae¹,

Baek²,

Ham³

et al. 2022

Preprint

View full text Add to dashboard Cite

show abstract

“…FPGAs and processor arrays have also been used as alternative architectures for JPEG [10] and Huffman decoding [11]. Different from GPU-based solutions these approaches focus on energy efficiency by designing optimized processing pipelines or bit-parallel architectures in hardware.…”

Section: Related Workmentioning

confidence: 99%

Accelerating JPEG Decompression on GPUs

Weißenberger¹,

Schmidt²

2021

Preprint

View full text Add to dashboard Cite

The JPEG compression format has been the standard for lossy image compression for over multiple decades, offering high compression rates at minor perceptual loss in image quality. For GPU-accelerated computer vision and deep learning tasks, such as the training of image classification models, efficient JPEG decoding is essential due to limitations in memory bandwidth. As many decoder implementations are CPU-based, decoded image data has to be transferred to accelerators like GPUs via interconnects such as PCI-E, implying decreased throughput rates. JPEG decoding therefore represents a considerable bottleneck in these pipelines. In contrast, efficiency could be vastly increased by utilizing a GPU-accelerated decoder. In this case, only compressed data needs to be transferred, as decoding will be handled by the accelerators. In order to design such a GPU-based decoder, the respective algorithms must be parallelized on a fine-grained level. However, parallel decoding of individual JPEG files represents a complex task. In this paper, we present an efficient method for JPEG image decompression on GPUs, which implements an important subset of the JPEG standard. The proposed algorithm evaluates codeword locations at arbitrary positions in the bitstream, thereby enabling parallel decompression of independent chunks. Our performance evaluation shows that on an A100 (V100) GPU our implementation can outperform the state-of-the-art implementations libjpeg-turbo (CPU) and nvJPEG (GPU) by a factor of up to 51 (34) and 8.0 (5.7). Furthermore, it achieves a speedup of up to 3.4 over nvJPEG accelerated with the dedicated hardware JPEG decoder on an A100.

show abstract