A 978GOPS/W Flexible Streaming Processor for RealTime Image Processing Applications in 22nm FDSOI stream up to 1024 cycles, buffering up to one line of HD frames without chaining FIFOs together. Unused SRAM blocks are placed in sleep mode for the full execution duration to minimize power consumption.

📋 论文概要

该论文提出了一种在22nm FDSOI工艺上实现的灵活流处理器，用于实时图像处理应用，能够以978GOPS/W的能效处理高达1024个周期的流数据。它通过可重构的处理器阵列和流架构，解决了传统图像处理加速器在灵活性和能效之间的权衡问题。

💡 主要创新点

• 提出了一种高度灵活的流处理器架构，支持多种图像处理内核（如Sobel滤波器和积分图像构建）的映射，无需重新设计硬件。
• 通过流水线处理和可配置的FIFO，实现了高效的流数据处理，支持高达1024个周期的连续流操作。
• 在22nm FDSOI工艺上实现了978GOPS/W的能效，显著优于传统通用处理器和固定功能加速器。

🏷 关键词

📄 原文摘要

of image processing compute kernels found in typical image processing pipelines, as listed in Fig. 2.2.1. Fig. 2.2.3 demonstrates how 3 of these kernels can be mapped on the presented fabric: horizontal and vertical Sobel filters and integral image construction. The extension from the first 2 kernels to larger or 2D filters is straightforward. For the third kernel, integral image construction, the first PE computes the running sum over a single line in the image and the PE with FIFO sums the current running sum with the result of the previous line. The reset of the values of the running sums at the start of a new line is implemented using a conditional operation using the guard control unit. Many of such basic image processing kernels can be concatenated in one long streaming pipeline across different interconnected PE units. We illustrate this for an optic flow application, using pyramidal Lucas-Kanade (LK) dense optical flow consisting of many

👥 作者与机构

Sander Smets1, Toon Goedemé1, Anurag Mittal2, Marian Verhelst1, Crucial to streaming fabrics is their ability to map and concatenate a wide range