High-level compute APIs and performance portability on Android

The versatility of heterogeneous platform had prompted it to become the mainstream of computer architecture design, especially for mobile computing. It is necessary to classify the strength and weakness among different programming models in heterogeneous system development to guide programmers to select suitable tools while implementing various software layers. In addition to the models, we will also offer two other angles: Hardware and runtime. The hardware side details heterogeneous compute devices. For the runtime angle, utilizing all available compute devices is a differentiator. Achieving device fusion and performance portability at run time have been active research areas recently. In this talk, we select two types of representative programming models, Google RenderScript and OpenCL, for comparison in order to fully illustrate their technical correlation and the implications to the platform architecture and system construction. We focus on the following aspects: language capability, performance, code complexity and multi-versioning, static runtime distribution and dynamic load balancing. RenderScript benefits application developers in terms of flexibility and hands-on programming convenience while OpenCL excels at wide adoption among hardware vendors and low-level runtime construction to maximize the system performance.