Steven Johnson

> LGTM, let me pull into google3 for some torture testing FYI, I'm failing on arm32-android builds with `ld.lld: error: undefined symbol: roundevenf`

OK, with that last change in place, i now have a fairly large number of failures that all seem to be variants on recursion-too-deep ending with SIGSEGV or similar; stand...

``` #0 0x000055555ada3e9a in Halide::Internal::Simplify::visit(Halide::Internal::Add const*, Halide::Internal::Simplify::ExprInfo*) () #1 0x000055555ad95ab8 in Halide::Expr Halide::Internal::VariadicVisitor::dispatch_expr(Halide::Internal::BaseExprNode const*, Halide::Internal::Simplify::ExprInfo*&) () #2 0x000055555ae0095c in Halide::Internal::Simplify::visit(Halide::Internal::Load const*, Halide::Internal::Simplify::ExprInfo*) () #3 0x000055555ad95bc8 in Halide::Expr Halide::Internal::VariadicVisitor::dispatch_expr(Halide::Internal::BaseExprNode const*, Halide::Internal::Simplify::ExprInfo*&)...

@vksnk -- there's an xtensa-related issue that is (hopefully) the final blocker here; LMK when you are back in the office. (Basically: we are missing a float32->uint32 vector reinterpret, or,...

OK, I think we're good to go.

I want to wait for LLVM on x86-32 to get fixed before we land this; a proposed fix is pending now, so hopefully only another day or so

This PR has been sitting here for a while -- should it land?

SGTM, but what would the equivalent signature be when making a call from Python? Maybe you could do something like `result = []` and have the caller insert the result...

Halide v14 doesn't support any version of LLVM newer than 14. It should be possible to cherry-pick the changes you need from Halide main, which does support LLVM 15/16.

An interesting wrinkle here is that a lot of the things you'd think need to be exposed (e.g. device_sync) are currently useless: the Python bindings don't actually provide `halide_buffer_t`, believe...