Metal.jl sincos intrinsic produces invalid LLVM IR on Julia 1.11+ (GPUCompiler 1.7.6+) #761
Description
Using the sincos intrinsic inside a Metal GPU kernel (KernelAbstractions.jl) fails at compile time with an InvalidIRError:
Environment:
[dde4c033] + Metal v1.9.1
[0c68f7d7] + GPUArrays v11.3.4
[61eb1bfa] + GPUCompiler v1.8.1This is a MWE:
using Metal, KernelAbstractions, Adapt
@kernel function sincos_kernel!(x)
i = @index(Global, Linear)
s, c = sincos(x[i])
x[i] = s + c
end
function sincos_caller!(x, backend)
sincos_kernel!(backend)(x; ndrange=length(x))
KernelAbstractions.synchronize(backend)
end
backend = MetalBackend()
x = rand(Float32, 100) |> adapt(backend)
sincos_caller!(x, backend) # <--- InvalidIRErrorRelated: JuliaHealth/KomaMRI.jl#716