Hardware-software co-design for image convolution on the Kria KV260
Abstract
Image convolution is a computationally intensive operation in embedded computer vision systems. This study aims to develop and evaluate a streaming field-programmable gate array (FPGA)–CPU hardware–software co-design for accelerating two-dimensional image convolution on the AMD Xilinx Kria KV260 platform. The accelerator was implemented using Vitis high-level synthesis (HLS) and Vitis Vision with Advanced eXtensible Interface 4 Stream (AXI4-Stream) interfaces for both image data and runtime-programmable filter coefficients. Two processing configurations were evaluated: XF NPPC1, processing one pixel per clock cycle, and XF NPPC8, processing eight pixels per cycle. Performance was compared with the OpenCV filter2D implementation executed on the ARM Cortex-A53 processor using grayscale and RGB images at resolutions up to 4K. Experimental results show that, for a 4K RGB 7×7 convolution, XF NPPC8 reduces end-to-end latency from 674 ms to 58.6 ms, providing an 11.50× speedup. For a 512×512 grayscale 3×3 workload, it achieves 990.1 frames per second (FPS) and 4.67 GOPS. System-level power consumption is 2.804 W for NPPC1 and 2.851 W for NPPC8, corresponding to energy efficiencies of approximately 0.52 and 1.64 GOPS/W, respectively. These results demonstrate that vector-level parallelism substantially improves throughput and energy efficiency while maintaining moderate system-level power consumption, supporting real-time high-resolution image-processing applications.
Keywords
All-programmable SoC; AXI4-Stream; Co-design; Image convolution; Image processing; Kria KV260
Full Text:
PDFDOI: https://doi.org/10.11591/eei.v15i6.12700
Refbacks
- There are currently no refbacks.

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
Bulletin of Electrical Engineering and Informatics (BEEI)
ISSN: 2089-3191
,
e-ISSN: 2302-9285
This journal is published by the
Institute of Advanced Engineering and Science (IAES)
.