Siu Kwan Lam
Siu Kwan Lam
Thank you for reporting and investigating into the problems. We were wondering why the `debug-only` stopped working. Since we will be bumping to a newer LLVM, we should make sure...
Re-posting from gitter https://gitter.im/numba/numba?at=6347db962a06f4566b341e00 It's worth looking at how JAX provide fft.
https://github.com/numba/numba/issues/9791 is likely another case
For your reference, I wrote a PoC custom extraction with custom cost model at https://github.com/sklam/prototypes_2025/commit/983b81d6e7f69a4444749039f605c670c96bd505 It is heavily inspired by Tensat (https://github.com/uwplse/tensat/blob/master/src/optimize.rs#L57C12-L57C25). The example expands `x ** 4` to `x...
I have been making progress with my POC. Here's a long notebook that goes through compiling a Python function, encoding it into egraph, custom extraction, and MLIR codegen: https://numba.github.io/sealir/demo_geglu_approx_mlir.html#extracting-an-optimized-representation I...
@saulshanabrook, I wonder if the builtin extraction can maintain the stable cost computation. I have tests for it at https://github.com/numba/sealir/blob/main/sealir/tests/test_cost_extraction.py. IMO, any way that can allow custom python-implemented extraction in...
Additionally, my team are going to investigate doing cost that are more than just a scalar. We want to investigate using some ML optimization technique to determine the cost-function. We...
One of the problem i found with tree cost is situation like `a ** 2` optimizing to `a * a`. It should not count cost of `a` twice. I captured...