SIMD machines — machines capable of evaluating the same instruction on several elements of data in parallel — are nowadays commonplace and diverse, be it in supercomputers, desktop computers or even mobile ones. Numerous tools and libraries can make use of that technology to speed up their computations, yet it could be argued that there is no library that provides a satisfying minimalistic, high-level and platform-agnostic interface for the C++ developer. In this talk, we shall present the upcoming Boost.SIMD library that is pending review, while explaining the challenges of the technology and how to adapt a series of common or not-so-common example problems to use it. In doing do, we aim to demonstrate the effectiveness of our library at easily speeding-up all kinds of algorithms, and at appropriately abstracting SIMD hardware so as not to exhibit any penalty. The design of Boost.SIMD is made so as to be as lightweight as possible; as a component of the larger numerical computation library NT2 — which uses it, along with SMP, MPI and GPGPU technologies, to build tables and matrices --, it is dedicated to only deal with SIMD. Therefore its main abstraction is that of the SIMD register, i.e. the base unit the SIMD processing unit manipulates; and albeit it provides a platform-agnostic and high-level interface, it is designed so that low-level issues can remain a primary concern to the user. Boost.SIMD also relies on the Boost.Proto DSEL framework so as to detect certain code patterns and map them to the most efficient solution. This is for example used for the fused multiplication/addition instructions, available on Altivec and on future generations of x86, but also for other things such as detection of values that are necessarily in a given range. Furthermore, the library embeds a powerful external dispatching mechanism which allows to fairly easily specialize a primitive for a given architecture or to create new targets or functions.

• published: 29 Oct 2013
• views: 32