high performance computing on graphics processing units: hgpu.org

JavaScript has seen tremendous growth in popularity driven by increasingly interactive web sites and sophisticated web interfaces. However, the performance of JavaScript continues to be a hurdle in using it for tasks that are computationally intensive, such as gaming, simulations, and visualization. JavaScript has also been slow to exploit the available parallelism on modern computers. Specifically, it is not currently easy to exploit GPGPUs within JavaScript. A part of the reason is that the low-level interface demanded for GPGPU programming is often not approachable by JavaScript programmers. In this paper, we present a novel approach that provides a mechanism to accelerate portions of JavaScript programs without requiring the programmers to learn new syntax or low-level APIs. We achieve that through an embedded DSL used to specify GPGPU computations. We have designed a JavaScript library, and an accompanying Firefox extension, that work together to compile the embedded DSL just-in-time using the LLVM backend for generating PTX. The compiled code is cached to minimize the compilation overhead. Our evaluation of the system using a micro-benchmark, two applications kernels, and an application benchmark demonstrates that our approach imposes minimal performance overhead, while providing an easy GPGPU programming interface to JavaScript programmers.