Usage of ::Int

[AntonReinhard]

I'm not sure if this is intentional, I assume it's to follow the general Julia default of Int, which is Int64 on most platforms. However, it also means I am unable to use Int32 when I specifically want to, and on GPUs it can actually make quite a difference especially for indexing or iterating.
As it stands, I cannot start a kernel with Int32 block sizes, for example:

  Expression: binning_vegas!(backend, bins_buffer, ndi_buffer, buffer, grid, normal_distribution)
  MethodError: no method matching KernelAbstractions.NDIteration.StaticSize(::Tuple{Int32, Int32, Int32})
  The type `KernelAbstractions.NDIteration.StaticSize` exists, but no method is defined for this combination of argument types when trying to construct it.
  
  Closest candidates are:
    KernelAbstractions.NDIteration.StaticSize(::Int64...)
     @ KernelAbstractions ~/.julia/packages/KernelAbstractions/aLmvZ/src/nditeration.jl:21
    KernelAbstractions.NDIteration.StaticSize(::Tuple{Vararg{Int64}})
     @ KernelAbstractions ~/.julia/packages/KernelAbstractions/aLmvZ/src/nditeration.jl:20
    KernelAbstractions.NDIteration.StaticSize(::Type{<:Tuple})
     @ KernelAbstractions ~/.julia/packages/KernelAbstractions/aLmvZ/src/nditeration.jl:22
  
  Stacktrace:
   [1] vegas_binning_kernel_batched!(dev::CPU, size::Tuple{Int32, Int32, Int32})
     @ Vegas ~/.julia/packages/KernelAbstractions/aLmvZ/src/macros.jl:68
   [2] binning_vegas!(backend::CPU, bins_buffer::Matrix{Float16}, ndi_buffer::Matrix{Int64}, buffer::VegasBatchBuffer{Float16, 4194304, 11, Matrix{Float16}, Vector{Float16}, Vector{Float16}}, grid::VegasGrid{Float16, 33, 11, Matrix{Float16}}, func::Function)
     @ Vegas ~/repos/Vegas.jl/src/binning.jl:169
   [3] macro expansion
     @ ~/repos/Vegas.jl/test/binning.jl:21 [inlined]
   [4] macro expansion
     @ ~/.julia/juliaup/julia-1.12.5+0.x64.linux.gnu/share/julia/stdlib/v1.12/Test/src/Test.jl:677 [inlined]
   [5] macro expansion
     @ ~/repos/Vegas.jl/test/binning.jl:21 [inlined]
   [6] macro expansion
     @ ~/.julia/juliaup/julia-1.12.5+0.x64.linux.gnu/share/julia/stdlib/v1.12/Test/src/Test.jl:1865 [inlined]
   [7] testsuite_binning(backend::CPU, el_type::Type{Float16}, nbins::Int64, dim::Int64)
     @ Main ~/repos/Vegas.jl/test/binning.jl:10

I think I could make a PR to make this more generic by using Integer instead of Int in most places.

[vchuravy]

A PR would be welcome! Note that development is on main, but we are also backporting PRs to release-0.9

[AntonReinhard]

Okay I have looked at it a little more closely. The particular problem I have is probably solvable on its own, but for a more general solution of currently hard-coded integer widths I think we would need to provide a lot of the intrinsics with an optional first type parameter. So things like @index(Global, NTuple) would need an optional @index(Int32, Global, NTuple).