Can someone point me to the documentation related to TensorFlow to XLA bridge using MLIR. I understand that it’s still a workin progress. I want to understand the stages(flow), where can i get more info on this?
There isn’t a nice concise one published yet (a summary one is in progress, but needs to go through a few iterations still). Currently most accurate result is enabling the logging and seeing the resultant passes.
From a high-level there are 2 stages: 1) rewrite for compilation/backend compiler, 2) legalization to backend compiler target and invocation. Given XLA, the first step handles the state management, complex types, replication representation, compile ops insertion etc and the 2nd primarily legalizing to HLO (well MHLO + CHLO + some optimizations + translate to HLO) & invoking XLA/device compilers