Back to All Events

PAST: Making use of SIMD Vectorisation to Improve Code Performance

WATCH HERE

Presenter: James Willis, SciNet

We are reaching the end of Moore's Law, the number of cores per chip is increasing and clock rates are peaking. Applications must be parallelised in order to reach peak theoretical performance of modern hardware. Alongside larger core counts, CPU vector registers have also increased in size. The latest chips support 512-bit vector registers operating with AVX512 instructions, which can give up to a 16x speed-up. In this talk we will introduce the concept of SIMD vectorisation and go through a simple example. We will also look at some optimisation techniques to improve vectorisation efficiency. Finally we will go over a few analysis tools that help identify potential code optimisations.

Previous
Previous
February 8

PAST: An Introduction to MPLAPACK, a Multi-precision Linear Algebra Library

Next
Next
February 22

PAST: Accelerated DataFrame with Dask-cuDF on multiple GPUs