Hacker News Reader

Check what LLVM is doing with the MLIR project.

https://blog.tensorflow.org/2019/04/mlir-new-intermediate-re...

https://llvm.org/devmtg/2019-04/talks.html

This is as good an overview as any: https://arxiv.org/pdf/1805.03441.pdf

In general we can afford to treat the problem as a search problem since our kernels are small enough. Here there is a lot of stuff that ML aided stochastic search is amazing at compared to normal compilers and even in some cases human hand optimized assembly (!).

You may enjoy the paper "Learning to Optimize Tensor Programs": https://arxiv.org/abs/1805.08166