Ben Vanik

Mali seems more right there - very small negative number :P Can we almost-equals compare instead? (e-6/e-7 is pretty tiny)

Awesome - that's helpful (and what I worried about)! This approach would let us work around that as each entry point would have its own shader module (if we needed),...

Fun post here showing MTLLibrary timings: https://github.com/gpuweb/gpuweb/issues/1064 (ideally we'd not have a flatbuffer for metal at all and just pack MTLLibrary's directly) WGSL will need the same approach as SPIR-V...

Do you know what the arith.constant is? (would help understand if it's a data type issue)

If so that's great - #8738 will add a new `-iree-flow-demote-i64-to-i32` flag that can be used for models that have i64 that should help!

#8738 may not fix tosa due to #8745 - it may need a corresponding `-iree-tosa-demote-i64-to-i32` on the tosa import path to the `-iree-mhlo-demote-i64-to-i32` on the mhlo path until it's fixed.

I don't know if "fixed" so much as worked-around - having i64 emulation on hot paths isn't great :) If we do start putting i64 emulation on hot paths we...

That's useful information! I was only trying to capture from an x86 host. Finding the right place to put some printfs that we can read back from the logs on...

I still have issues loading those android traces - I think I tracked it down to something that looked like undefined behavior somewhere in either the recording of string tables...

Hi @anthonycanino1! (sorry for the delay - was on vacation) scf.for has some canonicalization patterns here: https://github.com/llvm/llvm-project/blob/4184018253e720b0f2449b2b83ce27fc682f8579/mlir/lib/Dialect/SCF/SCF.cpp#L841-L845 The idea would be to add some new ones that perform the range...