mapreduce fails if init has different type

[maleadt]

julia> mapreduce((x) -> (x+1), +, rand(2,2); init=0)
5.16997158723571

julia> using CuArrays

julia> mapreduce((x) -> (x+1), +, cu(rand(2,2)); init=0)
ERROR: InvalidIRError: compiling partial_mapreduce_grid(var"#5#6", typeof(+), CUDAnative.CuDeviceArray{Float32,2,CUDAnative.AS.Global}, CUDAnative.CuDeviceArray{Int64,3,CUDAnative.AS.Global}, Int64, CartesianIndices{2,Tuple{Base.OneTo{Int64},Base.OneTo{Int64}}}, CartesianIndices{2,Tuple{Base.OneTo{Int64},Base.OneTo{Int64}}}, Int64, Val{true}) resulted in invalid LLVM IR
Reason: unsupported dynamic function invocation (call to reduce_block)
Stacktrace:
 [1] partial_mapreduce_grid at /home/tim/Julia/pkg/CuArrays/src/mapreduce.jl:122

[kshyatt]

Is this actually an init has different type issue, or an order-of-arguments issue? The problem I see now is that the float output of f is being converted into an Int, which of course doesn't work, but if I do:

julia> mapreduce((x) -> (x+1), +, cu(rand(Int, 2,2)); init=0.0)
1.5544613637878622e19

Everything's fine. So I think maybe the "root" here is we need a promote_type on init or something? A bit unsure... maybe a better way to put it is see if there is a way to infer the output array type from f, op, and init before allocating it?