[RFC] Changing the promotion type of Integer and Fixed to a float type.

[kimikage]

I was considering introducing checked_mul for Issue #78. However, I noticed that there are differences in promotion rules between Fixed and Normed.

julia> 2 * 0.5N2f14 # promotion to Float32
1.000061f0

julia> 2 * 0.5Q1f14 # promotion to Q1f14 (issue #78)
ERROR: ArgumentError: Fixed{Int16,14} is a 16-bit type representing 65536 values from -2.0 to 1.99994; cannot represent 2

Perhaps this is due to the current (historical) implementation where Fixed is signed and Normed is unsigned, but IMO, it seems practical to promote them to float types, and not to add ::Integer * ::FixedPoint.

[kimikage]

The rules for AbstractFloat are also subtly different. I think the Normed's rule is better.

FixedPointNumbers.jl/src/fixed.jl

Line 197 in 7c90fb6

promote_rule(::Type{Fixed{T,f}}, ::Type{TF}) where {T,f,TF <: AbstractFloat} = TF

FixedPointNumbers.jl/src/normed.jl

Line 310 in 7c90fb6

promote_rule(::Type{T}, ::Type{Tf}) where {T <: Normed,Tf <: AbstractFloat} = promote_type(floattype(T), Tf)

[kimikage]

If we decide to commonize the promotion rules, i.e. both Fixed and Normed are promoted to a float type for Integer, it seems to be "neutral" and practical that the promotion rule between Fixed and Normed returns a float types. (cf. #126)

[timholy]

Yes, we should go with the one used for Normed.

julia> x = rand(N0f32)
0.3137734384N0f32

julia> promote(x, Float16(1.0))
(0.3137734384075211, 1.0)

julia> x = rand(Q1f30)
-1.7184696067Q1f30

julia> promote(x, Float16(1.0))
(Float16(-1.719), Float16(1.0))

[kimikage]

I also want to commonize the following rules with #192 (comment).

FixedPointNumbers.jl/src/fixed.jl

Lines 200 to 213 in 0446a89

	@generated function promote_rule(::Type{Fixed{T1,f1}}, ::Type{Fixed{T2,f2}}) where {T1,T2,f1,f2}
	f = max(f1, f2) # ensure we have enough precision
	T = promote_type(T1, T2)
	# make sure we have enough integer bits
	i1, i2 = bitwidth(T1)-f1, bitwidth(T2)-f2 # number of integer bits for each
	i = bitwidth(T)-f
	while i < max(i1, i2)
	Tw = widen1(T)
	T == Tw && break
	T = Tw
	i = bitwidth(T)-f
	end
	:(Fixed{$T,$f})
	end

Also, I'm thinking of removing the @generated. It seems that the promotion result can be inferred by replacing the while loop with recursion.

Please note that this is a "general purpose" promotion rule and we can design the promotion rules for the arithmetic operations separately.

[timholy]

Yes, in general my view is that promote is not entirely adequate, because "promote for what?" is a relevant question. I've personally decided that promote really means "promote for cat", meaning it computes a single concrete type used as the eltype of a container that can store both objects.

From this standpoint, even promote among Nmfn types with different n should promote to floattype:

julia> promote_type(N0f8, N1f7) === N8f8
true

julia> x = rand(N0f8)
0.651N0f8

julia> float(x) == float(convert(N8f8, x))
true

julia> x = rand(N1f7)
1.661N1f7

julia> float(x) == float(convert(N8f8, x))
false

However, different Qmfn types can promote to the one with highest f (given adequate integer widening).

[kimikage]

Even in Normed, there are cases where conversion can be performed fast and losslessly, e.g. N0f8 to N0f16.
(In that sense, the comparison on float is not fair.)

Integers as promotion targets are often Ints i.e. with 32-bit or greater width, whereas FixedPoint is usually 8-bit or 16-bit. We're avoiding Float16, so I think the promotion to floattype is often overkill.

I prefer to apply the same promotion policy for the Fixed types and the Normed types.

[kimikage]

WIP: master...kimikage:promotion_rule
This requires nbitsint, which is going to be introduced in PR #189, to support mixing signed and unsigned (NOT Fixed and Normed)