Practical data-centric optimization of C/C++ applications

Abstract

C is a programming language used everywhere and in a big part of scientific codes. With the increase of highly parallel computing machines the code needs to be adapted to be able to use the full power available. In my thesis I will expand a compiler frontend to be able to transform C code into a data centric representation that will intern be able to identify data dependencies automatically and hence produce highly parallel code. There were a lot of challenges because pointers and structs are hard to adapt to a data centric paradigm. In the end I was able to find ways to work around the imposed limitations and expand the compiler front end to support a bigger - and more general - C language subset. On the tested benchmarks the performance of the automatic parallelization was on par with the handmade one.