Optimização e Geração de código OpenCL para computação embutida

Abstract

High-performance computing is increasingly multi-core. As such, sequencial computations must be replaced by their parallel counterparts for the hardware potential to be achieved. Programmers typically develop and validate sequential applications using high-level languages such as MATLAB and later manually migrate to lower-level parallel languages such as CUDA and OpenCL. This process is time consuming and error prone, so automatic source-to-source approaches are highly desirable, even when there is a considerable slow-down compared to the manual porting process. We present a tool capable of processing MATLAB and output the equivalent OpenCL and C wrapper code, based on an existing framework called MATISSE, which is already capable of generating C code from MATLAB source code. OpenCL generation is not entirely automated since it relies on the manual insertion of directives to guide the compilation. Future work could reduce this effort, since some directives are redundant and could be inferred by the compiler and others could be automatically injected into the code by Aspect-Oriented languages and tools.