Ben Vanik

Ahh I was wrong, there is a tiny difference: Release version: note the capitalized drive specifier in the `Formatting document` line (`D`): ``` LSP: (received) cpptools/formatDocument: file:///d%3A/Dev/iree/bugtest.c (id: 61) LSP:...

Thank you for the speedy response!

probably worth a quick sync - RE indirect: we don't really support them today anywhere else and disable optimizations in their presence, so that's fine to do here

Love what it's doing, but terrified by the PackOp and static data layout coupling. We really are going to need to stop using PackOp at the global opt level -...

Yeah, I've got some ideas - TLDR is encodings should be SSA values associated with tensors just like we do with dynamic dimensions, and just like we propagate/fold/inline dynamic dimensions...

Definitely don't need to peer inside executables - this is all feed-forward (symbolic encodings are assigned and kept consistent all the way until we know what devices they are used...

The encoding should be propagated before then. Think of encoding changes like casts or reshapes. I think w.r.t. this PR that's the interesting thing it's doing we may be able...

Yep - details are what we have to work through - it won't be possible to fully articulate them without actually solving the problem :)

sorry for the delay - I looked last week but didn't think it was up to date, as we'd chatted about separating this from tracing entirely and it looks like...

just wanted to note that @ScottTodd is bringing up the onnx test suite and there are a significant number of failures around conv ops - not sure if it's somewhere...