Data layout transformation using LLVM/Clang

Abstract

Modern systems contain a variety of data-parallel compute components, such as CPU vector units and GPU cores. Addressing these components requires a restructuring of application data so that it can be efficiently loaded from memory. In this paper, we present a backwards-compatible compiler-based data layout transformation technique that does not require preprocessing or runtime marshalling. Using our proposed extension, experimental results demonstrate a significant improvement over using an array-of-structures and can achieve near-performance-parity with manual transformation.