1.12

[christiangnrd]

I wasn't comfortable force-pushing in #385 so I opened this one.

[simeonschaub]

Unfortunately, this is currently blocked by #384 and potentially other issues. I thought I could work around it using the khronos backend, but looks like that is broken in other ways

[simeonschaub]

Suggested change

	let BitInteger64 = Union{Int64, UInt64}
	@device_override function Base.checkbounds(::Type{Bool}, v::StepRange{<:BitInteger64, <:BitInteger64}, i::BitInteger64)
	@inline
	return checkindex(Bool, eachindex(IndexLinear(), v), i)
	end
	end
	BitInteger64 = Union{Int64, UInt64}
	@device_override function Base.checkbounds(::Type{Bool}, v::StepRange{<:BitInteger64, <:BitInteger64}, i::BitInteger64)
	@inline
	return checkindex(Bool, eachindex(IndexLinear(), v), i)
	end

I don't think we can use let here and @device_override global function is not recognized by the macro. With this change and JuliaPackaging/Yggdrasil#12341 all tests are passing for me locally on 1.12

[christiangnrd]

I’ll make the change and revert my temporary test-reducing measures shortly.

Should we also limit POCL to 7.0 for the main package instead of just tests until someone figures out what’s causing the new test failures?

[simeonschaub]

I don't think we need to restrict compatibility for the entire package. The issue is incomplete Float16 support, which we didnt test previously because the jlls had Float16 support disabled

[simeonschaub]

We don't want to disable Float16 tests for the local pocl builds though, those should actually pass

[christiangnrd]

Tests are still failing which is what I was testing

[simeonschaub]

The PoCL local ones passed, e.g. https://github.com/JuliaGPU/OpenCL.jl/actions/runs/18655258708/job/53182708923

[christiangnrd]

That's only on 1.10 and 1.11 though, not 1.12

[simeonschaub]

Yes, but the 1.12 failures are not due to Float16, they're #384 as I wrote earlier. Once the SPIRV-LLVM-Backend jll has been registered we can retrigger CI

[christiangnrd]

Ok I understand now thank you for bearing with me. I'll push a change to skip float16 only on pocl_jll and to revert to using spirv_llvm_translator as soon as it's registered

[simeonschaub]

Sorry, I meant SPIRV-LLVM-Backend, not translator, so that part is already correct. I was just experimenting with the khronos translator, but it's still pretty broken atm

[christiangnrd]

@simeonschaub I just noticed that the local builds don't report Float16 support.

From the versioninfo():

 │  - Portable Computing Language
│    OpenCL 3.0, PoCL 7.2-pre main-0-gea1824d  Linux, Debug+Asserts, RELOC, SPIR-V, LLVM 19.1.7jl, SLEEF, DISTRO, POCL_DEBUG
└    · cpu-neoverse-n2-0xd49 (svm:c+f, usm:h+d, bda, fp64, il)

[simeonschaub]

Ah, that's arm though, right? I believe only the x64 builds have Float16 support

[christiangnrd]

Should we revert JuliaGPU/pocl@afa0682 and release pocl_jll v7.1.0+1?

[simeonschaub]

That commit is harmless, the one we might want to revert is JuliaGPU/pocl@893558a. But yes, that might be the best way forward

Edit: opened JuliaPackaging/Yggdrasil#12355

[christiangnrd]

Oops! I meant to link to the one you mentioned.

[christiangnrd]

Okay I think that when JuliaPackaging/Yggdrasil#12355 is registered and my suggestion reverting the check is applied, this will be ready to merge.

the only question left is do we keep testing both 1.11 and 1.12 or only 1.12? Like in the original 1.12 pr?

[simeonschaub]

the only question left is do we keep testing both 1.11 and 1.12 or only 1.12? Like in the original 1.12 pr?

We already have way too many CI jobs in my opinion. I think we can just drop 1.11 since it's not an officially supported release anymore

[christiangnrd]

I think the 1.12 macos runners just need more time.

[christiangnrd]

There seems to be a huge performance regression with ARM macOS on 1.12.
ARM macOS has some performance issues but unrelated to 1.12

[simeonschaub]

I think we can go ahead with this