Implicit Data Layout Optimization for Portable Parallel Programming in C++.

The programming process for modern parallel processors including multi-core CPUs and many-core GPUs (Graphics Processing Units) represents a significant challenge for application developers. We propose to use the widely-popular programming language C++ for parallel programming in a portable way, allowing the same program to be run on different target architectures. In this paper we extend our framework PACXX (Programming Accelerators in C++) with an additional compilation pass which simplifies data management for the programmer and makes the programming process less error-prone. These changes result in a significant reduction of execution stalls caused by memory throttling. We describe the implementation of the new data layout optimization and we report experimental results that confirm the advantages of our approach.