An embedded domain specific language for general purpose vectorization

Portable SIMD code generation is an open problem in modern High Performance Computing systems. Performance portability can already be achieved, however it might fail when user-framework interaction is required. Of all portable vectorization techniques, explicit vectorization, using wrapper-class libraries, is proven to achieve the fastest performance, however it does not exploit optimization opportunities outside the simplest algebraic primitives. A more advanced language is therefore required, but the design of a new independent language is not feasible due to its high costs. This work describes an Embedded Domain Specific Language for solving generalized 1-D vectorization problems. The language is implemented using C++ as a host language and published as a lightweight library. By decoupling expression creation from evaluation a wider range of problems can be solved, without sacrificing runtime efficiency. In this paper we discuss design patterns necessary, but not limited, to efficient EDSL implementation. We also study specific scenarios in which a language-based interface can surpass procedural interfaces in both efficiency, portability and ease of use. In particular we demonstrate higher performance when compared with equivalent BLAS Level 1 routines.