Compute for mobile devices

Abstract

For modern SoCs used in mobile devices, it is vital to focus on the processing efficiency through leveraging a heterogeneous potential of the architecture. In this tutorial, we offer a hands-on experience with existing APIs for accelerating compute-intensive portions of a mobile application.Specifically, as a first step we introduce the essentials of the most popular general Compute APIs available in the mobile domain, including RenderScript*, OpenCL*, GLES pixel and recently compute shaders (plus quick comparison to more vendor-specific APIs like CUDA* and Metal*). We continue with medium-complexity topics like example API-specific performance tricks in action.Finally, we touch advanced aspects like tools-assisted performance analysis. General focus is on the changes required for the typical user code to leverage good GPU acceleration. Main section: Deep-dive with the APIsIn parallel to the basic how-to micro-lessons on each API we describe important design aspects behind the API. This includes view of memory spaces between CPUs and GPUs, partition to host and devices in the programming model, etc. We also cover APIs interoperability topics like RenderScript and OpenCL interop with GLES.Also throughout the tutorial, we are introducing data-driven insights into the current SoC architectures, with particular focus on difference with conventional discrete/integrated GPUs. The focus is on the performance tips that community learned so far. During the tutorial we also keep spotting the differences/similarities of the APIs and the performance implications, concluding with performance tips that span multiple APIs and target devices.