Non-uniform structors of arrays in numba

CEMeNT-PSAAP / MCDC

MC/DC: Monte Carlo Dynamic Code

https://mcdc.readthedocs.io/en/latest/

BSD 3-Clause "New" or "Revised" License

22 stars 22 forks source link

Non-uniform structors of arrays in numba #158

Open jpmorgan98 opened 7 months ago

jpmorgan98 commented 7 months ago

Optimizations for how cross section data, tally meshes, and cells are stored. Right now with the np.ndarry the size of allocated memory for each item is that of the largest array and the same for all others. numba.jitclass is a remedy for this but is not GPU operable. Some initial ideas are

numba.jitclass with a puller function for GPU runs to nd.arrays
An offset scheme where data is stored as a single dim array with offsets
Use of other c-type data structures like pytorch or cupy

jpmorgan98 commented 7 months ago

@ilhamv how urgent is you're need to move towards jitclass?

ilhamv commented 7 months ago

@jpmorgan98 Per our discussion with Braxton yesterday, it sounds like it is easier for the GPU-Numba work, if we stick with the numpy structured array than if we move to use Jitclass, at least for now, perhaps until the first working GPU merge. But what do you think, @braxtoncuneo ?