GPU VSIPL: Core and Beyond

GPU VSIPL is an implementation of the Vector Signal Image Processing Library (VSIPL) and provides a librarybased solution to developing GPU-accelerated numerical applications. GPU VSIPL now supports much of the VSIPL Core Profile as well as element-wise matrix operators corresponding to required vector operators. This is nearly a five-fold increase in function count since GPU VSIPL was first introduced and includes numerous high-level numerical procedures. GPU VSIPL is implemented by optimized kernels written in the CUDA programming language. It provides support for real and complex matrix and vector operations; signal processing operations such as FIR, FFT, correlation, and histogram; and linear algebra operations such as matrix product and QR decomposition. GPU VSIPL achieves performance comparable to CUBLAS without encumbering the developer with CUDA-specific function calls and is link-compatible with existing VSIPL applications.