NEON Vectorization Workshop

We introduce vectorization with NEON ground up from the most basic concept up to very advanced vectorization topics.

Workshop agenda:

• A short introduction to vectorization
• Introduction to NEON intrinsics
• Advanced NEON intrinsics
• Basic vectorization patterns – vectorizing for loops, foor loops with early exit, while loops and convergence loops
• Common vectorization patterns – vectorizing loops with conditions, conditional counting, loops with structs and matrix transposition.
• Vectorization inhibitors – learn to detect and remove obstacles that hinder efficient vectorization
• Vectorization types according to data access pattern – there are several ways to do vectorization, here we investigate inner-loop vectorization, outer-loop vectorization.
• Advanced vectorization patterns – we talk about how to vectorize copy_if, trees and lookup tables.
• Memory performance – improve the performance of your vectorized code by better using the memory subsystem.
• Peak performance – reach peak software performance by breaking instruction dependencies, avoiding register spills and cleverly using everything hardware has to offer.

Hardware/Software requirements
The workshop can be done on real hardware or on an emulator.

For doing the NEON workshop on real hardware, you will need:

An ARM 64-bit CPU with NEON extensions. These can be:

• Any embedded CPU with ARM64 and Linux, e.g. Raspberry Pi 3 or later supports this. On a Linux ARM system, you can check by running: lscpu | grep -e asimd -e aarch64. The output of this command should be:

Architecture: aarch64
Flags: fp asimd evtstrm crc32 cpuid,

• A MacOS device with Apple Silicon M1 or later.

You should have g++ installed for Linux or XCode for Mac.

For doing the NEON workshop on an emulator, you will need:

• An x86-64 system with either Ubuntu based Linux OR Ubuntu running on Windows Services for Linux (WSL)

qemu-aarch64 for emulating the system (available in the Ubuntu repositories through qemu-user package
aarch64-linux-gnu-g++ compiler (available in the Ubuntu repositories through g++-aarch64-linux-gnu package)