Jeremy Kun

ok so we actually have a slight problem here in that the manylinux specification forbids linking against openmp (linking against any shared lib not in this list implies noncompliance: https://peps.python.org/pep-0599/#the-manylinux2014-policy)...

Specifically `cibuildwheel` builds and then fails at ```bash Repairing wheel... + sh -c 'auditwheel repair -w /tmp/cibuildwheel/repaired_wheel /tmp/cibuildwheel/built_wheel/heir_py-0.0.2-cp311-abi3-linux_x86_64.whl' INFO:auditwheel.main_repair:Repairing heir_py-0.0.2-cp311-abi3-linux_x86_64.whl auditwheel: error: cannot repair "/tmp/cibuildwheel/built_wheel/heir_py-0.0.2-cp311-abi3-linux_x86_64.whl" to "manylinux_2_28_x86_64" ABI because of...

The errors like `libc.so.6 offending versions: GLIBC_2.32` I can fix by bumping to a newer `manylinux` container version, but the libgomp error would not be fixed by this.

Just to unblock releases at this point, I'm going to disable openmp in the heir-py build and file an issue.

Another option: https://github.com/cerisier/toolchains_llvm_bootstrapped

This is the first I've heard of the `top` dialect... I tried to find a source for this, and I think it's from https://github.com/sophgo/tpu-mlir ? Maybe there's a binary from...

Besides going onnx->iree I think onnx also directly supports converting onnx dialect to stableHLO, and stablehlo-opt supports lowering to linalg, and we can take it from there. https://github.com/onnx/onnx-mlir/tree/main/src/Tools/onnx-mlir-opt

FYI, this diff sufficed for me to get it to work for my website. Happy to send it as a PR if it seems OK ```diff diff --git a/static/webmention.js b/static/webmention.js...

A _perfect_ solution would be to have, for each backend target, a relative latency cost for each scheme operation that is derived from measured microbenchmarks of the backend target. I.e.,...

This appears to be a script that can be used to generate a cost model for Lattigo on CPU: https://github.com/applexi/lattigo/blob/main/benchmarks/cost_model_gen.go