high performance computing on graphics processing units: hgpu.org

This document intends to provide a detailed guidance on how to optimize OpenCL programs with Adreno GPUs. A good amount of information has been provided to help developers understand the OpenCL fundamentals and Adreno architectures, and most importantly, master OpenCL optimization techniques. OpenCL optimization is often challenging and requires a lot of trial and error. As each vendor may have its own best practices of doing the same task, it is important to read through and have in-depth understanding of the guide and practices for Adreno GPUs. Many factors that look minor could have significant performance impacts. These are unfortunately not easy to tackle without hands-on exercise and practices. Due to time constraints and other factors, some topics are not covered. Adreno GPUs support a lot of extensions that can significantly boost performance and add extra functionalities. For instance, recent Adreno GPUs support some proprietary image formats which raw image and video data captured from image signal processor (ISP) can be compressed to for direct and efficient processing. This could save some manual handling as well as improve bandwidth usage. Future releases of this document will include more topics.