PDCON:Conference/GPU audio signals processing in Pd, and PDCUDA, an implementation with the CUDA runtime API

An implementation named PDCUDA is presented for use of graphics processing unit (GPU) general programming capability for audio signals processing with Pure Data (Pd) and the CUDA runtime application programmers interface (API). The goal of PDCUDA is to make uniformly efficient the development of CUDA based externals and their usage in Pd. In PDCUDA, the Pd source code is patched to enable the creation of CUDA digital signals processing (DSP) contexts, in which signal memory is allocated in GPU memory and DSP operations are performed by the GPU. This capability is made accessible to users in Pd's graphical patching environment by modifying the canvas class. PDCUDA adds a new canvas creation method for the symbol “cucanvas” which identifies that the underlying DSP context for that canvas uses CUDA. Memory transfer between host and GPU explicitly occurs between signals in a non-root canvas and its parent canvas. The Pd source code for DSP scheduling is reviewed for existing methods of optimizing memory access and avoiding latency. PDCUDA provides shared/static libraries of functions and a header file comprising an API for externals developers.