Test for min / max broadcasting issue

[rkierulf]

I've run into an issue broadcasting min and max together over a Metal array:

using Metal
A = Float32[1,2,3,4,5]
B = convert(MtlArray,A)
t_start, t_end = Float32(1.0), Float32(4.0)
min.(max.((A .- t_start) ./ (t_end - t_start), zero(Float32)), one(Float32))
min.(max.((B .- t_start) ./ (t_end - t_start), zero(Float32)), one(Float32))

The last two lines produce different results, with the results for the CPU array A being correct and the GPU array B being incorrect. For the GPU result, it seems to produce the same results as the line below, interestingly:

max.((B .- t_start) ./ (t_end - t_start), one(Float32))

I'm seeing this on Julia 1.10.4 with a Mac M1, and I've confirmed this issue to be the cause of this failing build: https://buildkite.com/julialang/komamri-dot-jl/builds/824 which passes for Julia 1.9, so it's possible this is a regression in Julia 1.10, but I haven't confirmed this to be the case. For our use case, we should be able to work around the issue by using a tmp array to separate the min / max broadcasts, or writing an actual kernel to do the equivalent of the failing operation.

Full version info:

Julia Version 1.10.4
Commit 48d4fd48430 (2024-06-04 10:41 UTC)
Build Info:
  Official https://julialang.org/ release
Platform Info:
  OS: macOS (arm64-apple-darwin22.4.0)
  CPU: 8 × Apple M1
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-15.0.7 (ORCJIT, apple-m1)
Threads: 1 default, 0 interactive, 1 GC (on 4 virtual cores)

[tgymnich]

Seems like the custom maximum / minimum implementation for metal in juliagpu/GPUCompiler is broken.

[christiangnrd]

Bisected to JuliaLang/julia#47814.

[maleadt]

Would probably be valuable to add this as a test though.

[cncastillo]

I just noticed that this was fixed in GPUCompiler.jl. Would pushing a PR bumping its version be helpful?

[christiangnrd]

@cncastillo The minimum supported Julia version was increased to 1.10 on master/main for both repos. I assume the next GPUCompiler version will be breaking and that's why it hasn't been bumped yet.