M. Kretz (Universität Frankfurt): Explicitly Data-Parallel Programming with C++
Wednesday, 20.05.2015 16:15 im Raum M5
Computing hardware is steadily increasing the attainable operations executed
per second. However, this increase is only accessible through proper
parallelization of our software. While the multi-core/multi-thread issue has
seen a lot of research and solutions in recent years, the intrinsic
parallelism per CPU core (SIMD) has widely been neglected. A look at current
hardware shows that our software should take SIMD execution seriously.
Starting from a quick introduction to data-parallelism, I will present how
this common software property can be translated to SIMD instructions and thus
a considerable speedup of the software. This leads to the programming problem:
Our programming languages are still unable to express synchronously data-
parallel (e.g. SIMD) execution. I will present the current plans of the C++
committee for solving the issue. Vector types, as one of the solutions, allow
explicit data-parallel programming via the type system. Since most developers
do not want to wait for another 6 years, I will focus on SIMD programming with
the Vc library, which is the basis for the C++ committee work. I will round up
the talk with examples, presenting applications and upcoming extensions of the
Vc library.
(Added 2015-05-24) Links:
Slides, Old Vc Project Page, New Vc Repository on Github
Angelegt am Thursday, 19.03.2015 20:11 von Christian Engwer
Geändert am Monday, 25.05.2015 17:24 von Jorrit Fahlke
[Edit | Vorlage]