AcceleratedKernels.cumsum runs endlessly when using pocl's cpu backend on Intel CPU

[VarLad]

I've been able to reproduce it on two Intel devices. Would be nice if someone could test on AMD.
Rusticl works just fine here.

Julia: 1.11.3
OpenCL: 0.10.5
pocl: pocl_jll.jl (v7.0.0+15)

[christiangnrd]

Can you add a MWE? I'll test it on all my devices if I can just copy-paste the code.

[VarLad]

@christiangnrd

using OpenCL, AcceleratedKernels, pocl_jll

x = CLArray([1, 2, 3, 4, 5])
y = AcceleratedKernels.cumsum(x)

[VarLad]

I just tried the above on Julia 1.12
Same issue persists

[christiangnrd]

Are you running this in a REPL?

[VarLad]

Yeah, in REPL and in a Pluto notebook

[christiangnrd]

Can you test my even lazier MWE to make sure it can reproduce the issue for you?

using Pkg; Pkg.activate(temp=true); Pkg.add(["OpenCL", "AcceleratedKernels", "pocl_jll"]); using OpenCL, AcceleratedKernels, pocl_jll; begin
    x = CLArray([1, 2, 3, 4, 5])
    y = AcceleratedKernels.cumsum(x)
end

[VarLad]

Yeap, still running, so I guess still the same issue.
@christiangnrd Also, if I replace pocl_jll with Rusticl_jll, things just work, so I'm assuming this is a pocl issue for now 🤔

[christiangnrd]

Does it hang if you add a semi-colon after the begin end block?
I think the issue is triggered when trying to display USMBackend-backed arrays.

[VarLad]

Does it hang if you add a semi-colon after the begin end block?

No, but if I do anything with y and try copying it back to a Julia backed array, it does hang.

I think the issue is triggered when trying to display USMBackend-backed arrays.

Can still reproduce it with:

using Pkg; Pkg.activate(temp=true); Pkg.add(["OpenCL", "AcceleratedKernels", "pocl_jll"]); using OpenCL, AcceleratedKernels, pocl_jll; begin
    x = CLArray{Int64, 1, OpenCL.cl.Buffer}([1, 2, 3, 4, 5])
    y = AcceleratedKernels.cumsum(x)
end

so thats unlikely.

I think this used to work for some version of pocl, so the best way to bisect is probably to try older versions.

[pjaaskel]

If you can, please create a small reproducer and we can test it with the new compiler branch as well.

[christiangnrd]

@VarLad It works for me when I set a a lower block_size value.

The following does not hang but it does seem to give incorrect results.

using Pkg; Pkg.activate(temp=true); Pkg.add(["OpenCL", "AcceleratedKernels", "pocl_jll"]); using OpenCL, AcceleratedKernels, pocl_jll; begin
    x = CLArray{Int64, 1, OpenCL.cl.Buffer}([1, 2, 3, 4, 5])
    y = AcceleratedKernels.cumsum(x; block_size=16)
end

[VarLad]

Still reproducible as of POCL 7.1

[christiangnrd]

Inspired by pocl/pocl#1971 (comment), I tried this by launching julia with POCL_WORK_GROUP_METHOD=cbs and I can no longer reproduce the hangs at a greater block size nor the incorrect responses at lower block sizes.

[VarLad]

Good find! :)