Silent type promotion when creating range using `a🅱c`

General Usage
1

I was meddling with ranges and found the following strange behaviour: suppose I want to create a range starting at 0, with step of 1 and length 10. If I want to work with Float32s, I’d type something like

r1 = 0.0f0:1.0f0:9.0f0

I then tried to use this within some KernelAbstraction GPU kernels and I kept getting errors.

Upon further investigation, I saw that

typeof(r1) = StepRangeLen{Float32, Float64, Float64, Int64}

so I got Float64s in the the ref and step fields of StepRangeLen, instead of the expected Float32.

If I explicitly define

r2 = StepRangeLen(0f0, 1f0, 10)

I get the expected

typeof(r2) = StepRangeLen{Float32, Float32, Float32, Int64}

and the errors I got from the GPU kernel go away.

My question is if this silent type promotion is intended, and, if so, what is the reason behind it.

2

What’s happening is more explicit with Float64:

julia> typeof(0:0.1:1)
StepRangeLen{Float64, Base.TwicePrecision{Float64}, Base.TwicePrecision{Float64}, Int64}

It’s a design choice, and certainly a debatable one at that. It’s done to avoid roundoff issues. Compare

julia> collect(0:0.1:1)
11-element Vector{Float64}:
 0.0
 0.1
 0.2
 0.3
 0.4
 0.5
 0.6
 0.7
 0.8
 0.9
 1.0

julia> collect(StepRangeLen(0,0.1,11))
11-element Vector{Float64}:
 0.0
 0.1
 0.2
 0.30000000000000004
 0.4
 0.5
 0.6000000000000001
 0.7000000000000001
 0.8
 0.9
 1.0

This issue appears cosmetic, but it can actually lead to significant ambiguity. Knowing that 0.1*3 == 0.30000000000000004 > 0.3, should last(0.0:0.1:0.3) be 2.0, 3.0, or 0.30000000000000004? The TwicePrecision step is the only way to give the user what they (usually) want while maintaining consistency.

But I will sympathize that I, too, find it a tedious to have to use the StepRangeLen constructor when I’m okay with a lower precision step. Especially because there is a performance cost.

3

You might be looking for LinRange(0f0, 9f0, 10), which is the low-tech cousin of StepRangeLen, with none of this higher-precision-intermediate cleverness.

2 Likes