Ferramenta de compilação para a geração de código para um runtime de offloading automático

Abstract

Heterogeneous computing is an area in computer science that has become of great interest both comercially and academically. This area encompasses our understanding of computer nodes formed by both accelerating devices, such as GPUs and multi core processors. Although popular, very few applications are able to use all of the computing power available in these heterogeneous platforms. With these questions in mind, an execution enviroment was developed to solve this problem. This enviroment takes as input blocks of code (functions) with two different versions: one version destined to the GPU and another version destined to the CPU. This enviroment is capable of controlling the threads in the CPU and use them to collect data to decide which device is best suited for a specific function inside of the application. In this project, we attempted to develop a compiler tool with the purpose of working with this execution enviroment. The tool is able to detect if a demarcated region of code is parallel, starting from a code originally sequencial. After that, the tool generates code for both CPU by means of function calls to the OpenMP library and GPU by means of CUDA. To show that the compiler tool is functional, we produced a set of conceptual proofs that show the step-by-step of the tool by explaining what happened during the compilation process, the transformations made during each step and finally the output generated by the enviroment, proving that the system created and developed here is viable for improving the efficiency of code with very little changes in the original source code.