Mismatch between input datatype and interpolated type for `Float32`

[hzarei4]

Suppose you have an array of

a = rand(Float32, (10, 11))

As you have mentioned in this commit, the eltype of the interpolate object

itp = interpolate(a, BSpline(Linear()))

is Float32. Then we want to calculate the value of

julia> typeof(itp(1.1, 1.1))
Float64

Then, interestingly

julia> typeof(itp(1.1f0, 1.1f0))
Float32

I think this is an issue with handling the datatypes. I suggest that the interpolation output type should be the input data a we feed the function with, NOT the input type of the itp.

[mkitti]

Thanks. That seems reasonable. Could you make a pull request?

[henry2004y]

I have also encountered this issue many times, and have been wondering what is the "correct" treatment of the output type.

[stevengj]

I think the current behavior is correct, and consistent with Julia's usual promotion rules. If you do math with a combination of Float32 and Float64 values, you get a Float64 value.

For example, if you define a function

f(x) = 1.0f0 + 2.0f0 * x

then you would expect to get a Float64 value from f(1.1)

Your single-precision itp object is basically just like this function.