Portable Support for Explicit Vectorisation in C
A. Šinkarovs and S.-B. Scholz, “Portable support for explicit vectorisation in C,” in 16th workshop on compilers for parallel computing (CPC’12), Jan. 2012.
Abstract
In pursuit of requirements of modern software, high-performance computing often offers a narrowly-tailored solutions that reduce portability of software. As one of such examples in this paper, we consider the situation with SIMD accelerators, whose importance seriously increased in the last decade. Firstly appeared in the early 90es, being oriented exclusively on graphics acceleration, nowadays SIMD CPU extensions are used in a variety of fields. The lack of standard and incompatibility of instruction sets through the different types of CPUs substantially increase the complexity of developing portable applications within the extensions. In this paper we present an abstraction layer implemented as a set of C language extensions within the GNU GCC compiler which provides an interface for SIMD vectors and operations independently from the architecture. First of all, these abstractions allow to exploit SIMD extensions of a CPU explicitly, which is useful when auto-vectoriser fails. Secondly, the abstractions are general enough to be mapped to any hardware supporting SIMD paradigms; hence the new abstractions could be considered as a step forward to a new C language standard.