CppCon 2016

Love writing blazing fast SIMD code on CPU? Tired of dealing with ugly intrinsics and clumsy SIMD float4 classes? Has your compiler's auto-vectorization ever stopped working, causing unpredictable performance regressions? Wish you could write efficient SIMD code without locking yourself into a specific instruction set, while still taking advantage of a range of hardware from old desktops to new Intel Xeon Phi rigs?

The solution is here, and it's called SPMD!

SPMD is an elegant parallel programming technique for writing SIMD code, which automates the tedious constructions normally required when using intrinsics or assembly, breaks free of ties to specific instruction sets, and still allows you to work at the granularity of SIMD vectors when necessary.

This talk will first review the underlying theory of SPMD, then introduce the ISPC programming language as an example of what SPMD code looks like. We'll also look at how SPMD is used on GPUs like AMD's GCN. Finally, we will see how SPMD code can be written in unextended C++.