Notes: GPU IRs
from a conversation with Daniel after https://sflean.group/
- pytorch -> triton -> tritonIR -> ptx -> sass
- or eager-mode: pytorch -> prebuilt cuda kernels
- jax -> jaxpr -> xla -> llvm ir -> ptx -> sass
- can’t have if statements in jax - branches are ?collapsed? at compiletime
- there is a keyword in jax that gets around this
- it lowers the branch to the corresponding function in xla