[RFC] What should the arithmetic within the FixedPointNumbers be

Specific Domains Visualization
, ,
12

Not quite so simple; on Julia 1.5 and 1.6 I get the same time from both of those, and both convert to Gray{Float32} as they go. To prevent mean from doing something sneaky, this may be better:

julia> function naivesum(A)
           s = zero(eltype(A))
           @inbounds @simd for a in A
               s += a
           end
           return s
       end
naivesum (generic function with 1 method)

julia> @btime naivesum($v1)
  3.035 μs (0 allocations: 0 bytes)
Gray{N0f8}(0.125)

julia> @btime naivesum($v2)
  16.653 μs (0 allocations: 0 bytes)
Gray{Float32}(131201.62f0)

As you can see, there’s actually a cost to promoting to Float32. Of course, some kind of promotion is necessary to get the right answer here.

EDIT: you can compare various strategies with

julia> function naivesum(::Type{T}, A) where T
           s = zero(T)
           @inbounds @simd for a in A
               s += a
           end
           return s
       end
naivesum (generic function with 2 methods)

julia> @btime naivesum(Gray{N24f8}, $v1)
  8.226 μs (0 allocations: 0 bytes)
Gray{N24f8}(1.30728e5)

julia> @btime naivesum(Gray{Float32}, $v1)
  13.343 μs (0 allocations: 0 bytes)
Gray{Float32}(130727.81f0)

which is one of the reasons I’ve wondered above promoting to a wider FPN type.

1 Like