Performance Efficiency of Image Processing Algorithms on GPU using GPU specific languages

Performance efficiency focuses on how an application’s performance scales relative to its theoretical peak on a hardware. The work examines porting two image processing algorithms, Gaussian Filter and Connected Components Labeling, across NVIDIA, AMD, and Intel GPUs using native kernels. We introduce a performance model to evaluate efficiency, noting that simply using peak throughput and memory bandwidth metrics is insufficient due to varying compute and memory demands across applications. We see that the CUDA implementation on an NVIDIA A10 GPU aligns more closely with the performance model than the hip implementation on AMD Mi210 or the DPC++ implementation on Intel Flex 170. This is attributed to the strong hardware-software co-design in CUDA. Additionally, non-CUDA implementations on NVIDIA GPUs underperform compared to CUDA, reiterating the need for GPU-native kernels and further research into optimizing generic software platforms for performance efficiency.