Implementing Performance Libraries on Graphics Hardware

We propose a simple method to implement floating-point vector math operations and matrix multiplication on graphics hardware, focusing on identification of details, in both software and hardware, which affect performance and ease of use. Before widespread adoption of the graphics processing unit (GPU) as another computation processor, we must address the need of application interfaces (APIs) that abstract away the details of the implementation. We focus on providing an interface to the hardware that utilizes high level interfaces that hide the specifics of implementing the functionality on the GPU, while maintaining performance. We then use this interface to implement non-negative matrix factorization, used for performing feature extraction, to demonstrate the strengths of the library when run on current graphics hardware.