summary.md

Technical Details

1. It should exclusively use OMP for threading
2. It should be written with OMP4.5 in mind, as well as being CUDA extensible

Considerations

One of the more interesting things about PIC for advanced architecrues is that the different aspects of the push require different (contradictory) optimizations.

The main particle_push has three main parts:

The particle move

Particle Properties Used (memory streams): all
Data layouts: i) AoS => Good, not great vectorized ii) SoA => Good, but many memory streams iii) AoSoA => Ideal

• If we vectorize AoS we need to do a transpose
• SoA doesn't really buy us anything in terms of memory streams here as we need all properties.
• Particle order isn't a concern here
• There isn't much burden on cache here as long as the next particle(s) are pre-fetched in time

The field stencil (read) and particle velocity

Particle Properties Used (memory streams): Most
Data layouts: i) AoS => Not great, doesn't need to use all the streams ii) SoA => Good, lets you split the streams out iii) AoSoA => Great

• The field stencil benefits significantly from having well ordered (on a cell basis) particles
• Cache use here is crucial because of the large, semi-nonobvious (prefetching), stencil
• If we can explicitly tell the compiler that groups of particle share properties, we get better re-use than good "accidental" cache reuse

The current accumulation stencil (write)

Particle Properties Used (memory streams): Momentum
Data layouts: i) AoS => Not great, doesn't need to use all the streams ii) SoA => Good, lets you split the streams out iii) AoSoA => Great

• If particles all write to the same cell, that gives good assumptions and makes it safe to do some writes
• If particles are highly disordered, this can cause big problems for the safety of the writes and often leads to atomics