Illegal memory access in GemmKernels on Julia 1.9

[thomasfaingnaert]

As requested: we're getting a CUDA error: an illegal memory access was encountered (code 700, ERROR_ILLEGAL_ADDRESS) in the benchmark suite of GemmKernels. I placed a reproducer at https://github.com/JuliaGPU/GemmKernels.jl/tree/tf/repro-illegal-memacces.

Interestingly, the illegal memory access only happens on Julia 1.9, and is fixed on 1.10:

$ git clone https://github.com/JuliaGPU/GemmKernels.jl --branch tf/repro-illegal-memacces
$ cd GemmKernels.jl/

$ julia-1.9.4 --project -e 'using Pkg; Pkg.instantiate(); include("repro.jl")'
ERROR: LoadError: CUDA error: an illegal memory access was encountered (code 700, ERROR_ILLEGAL_ADDRESS)
# ...

$ julia-1.10.0 --project -e 'using Pkg; Pkg.instantiate(); include("repro.jl")' 
# works

ref JuliaGPU/GemmKernels.jl#186 (comment)