CppCon 2016

Parallel programming is a hot topic, and everybody knows that multicore processors and GPUs can be used to speed up calculations. What many people don't realize, however, is that CPUs provide another way to exploit parallelism -- one that predates recent multicore processors, has less overhead, requires no runtime scheduler, and can be used in combination with multicore processing to achieve even more speedup. It's called vector parallelism, and the hardware that implements it goes by brand names like SSE, AVX, NEON, and Altivec. If your parallel program does not use vectorization, you could be leaving a factor of 4 to 16 in performance on the floor.

In some ways, Vector programming is easier than thread-based parallel programming because it provides ordering guarantees that more closely resemble serial programming. Without an intuitive framework by which to interpret them, the ordering rules can be confusing, however, and restrictions on vector code that don't apply to thread-parallel code must be kept in mind. In this talk, we'll introduce you to the common elements of most vector hardware, show what kind of C++ code can be automatically vectorized by a smart compiler, and talk about programmer-specified vectorization in OpenMP as well as proposals making their way through the C++ standards committee. You'll understand the rules of vectorization, so that you can begin to take advantage of the vector units already in your CPU.

A basic understanding of C++11 lambda expressions is helpful.