Stream: helpdesk (published)

Topic: ✔ Division and unwanted `Float64`

Alec (Mar 21 2025 at 13:57):

I'm dabbling with some GPU kernel programming and hitting some issues with Float64s that crop up in the math. Specifically, I've got some integers and when I divide them I get a Float64. E.g.

Julia> /(Int32(4),Int32(5)) |> typeof
Float64

Any tips for handling this, such as an alternative to / that will keep the results as Float32? Do I just have to wrap every / operation with Float32(...)?

Cédric Belmant (Mar 21 2025 at 14:47):

You might want to convert these integers as Float32 before perfoming the division; that is what the method for /(::Int32, ::Int32)/ does but it uses float to convert them to Float64.

Neven Sajko (Mar 21 2025 at 14:47):

function d(a::Integer, b::Integer)
    Float32(a) / b
end
function d(a, b)
    a / b
end

Gunnar Farnebäck (Mar 21 2025 at 15:00):

Alec said:

Any tips for handling this, such as an alternative to / that will keep the results as Float32?

Keeping the result as Float32 sounds inaccurate if neither argument was Float32 to begin with. Sometimes you can get away with reordering your computations, e.g. if x is Float32 change 4 / 5 * x to 4 * x / 5 but usually it's easiest to add explicit casts where needed or use Float32 literals, e.g. 4f0 / 5 * x.

Gunnar Farnebäck (Mar 21 2025 at 15:10):

You also have the option to express the divisions as rationals, e.g. 4 // 5 * x. In the case of literals this is optimized to just a Float32 multiplication.

julia> f(x) = 4 // 5 * x
f (generic function with 1 method)

julia> @code_typed f(1.0f0)
CodeInfo(
1 ─ %1 = Base.mul_float(0.8f0, x)::Float32
└──      return %1
) => Float32

Daniel Wennberg (Mar 21 2025 at 16:05):

Gunnar Farnebäck said:

You also have the option to express the divisions as rationals, e.g. 4 // 5 * x

This :point_of_information: is the general solution to avoid spurious promotion, at least as long as you're only doing basic arithmetic. Related to this performance tip from the manual: https://docs.julialang.org/en/v1/manual/style-guide/#Avoid-using-floats-for-numeric-literals-in-generic-code-when-possible.

There's also zero, one, and oneunit, which are helpful if you're doing more than plain arithmetic and want to keep your code generic. For example, if x is a float of some type and I want to do x * sqrt(3), I'll write x * sqrt(3one(x)) to avoid spurious promotion. But for GPU kernel programming, you're probably not all that concerned about keeping your code generic.

Alec (Mar 21 2025 at 20:38):

Thanks all, very helpful!

Notification Bot (Mar 21 2025 at 20:38):

Alec has marked this topic as resolved.

Last updated: Apr 25 2025 at 04:43 UTC