Julia has a system for promoting arguments of mathematical operators to a common type, which has been mentioned in various other sections, including Integers and Floating-Point Numbers, Mathematical Operations and Elementary Functions, Types, and Methods. In this section, we explain how this promotion system works, as well as how to extend it to new types and apply it to functions besides built-in mathematical operators. Traditionally, programming languages fall into two camps with respect to promotion of arithmetic arguments:
+
, -
, *
, and /
, are automatically promoted to a common type to produce the expected results. C, Java, Perl, and Python, to name a few, all correctly compute the sum 1 + 1.5
as the floating-point value 2.5
, even though one of the operands to +
is an integer. These systems are convenient and designed carefully enough that they are generally all-but-invisible to the programmer: hardly anyone consciously thinks of this promotion taking place when writing such an expression, but compilers and interpreters must perform conversion before addition since integers and floating-point values cannot be added as-is. Complex rules for such automatic conversions are thus inevitably part of specifications and implementations for such languages.1 + 1.5
would be a compilation error in both Ada and ML. Instead one must write real(1) + 1.5
, explicitly converting the integer 1
to a floating-point value before performing addition. Explicit conversion everywhere is so inconvenient, however, that even Ada has some degree of automatic conversion: integer literals are promoted to the expected integer type automatically, and floating-point literals are similarly promoted to appropriate floating-point types.In a sense, Julia falls into the “no automatic promotion” category: mathematical operators are just functions with special syntax, and the arguments of functions are never automatically converted. However, one may observe that applying mathematical operations to a wide variety of mixed argument types is just an extreme case of polymorphic multiple dispatch — something which Julia’s dispatch and type systems are particularly well-suited to handle. “Automatic” promotion of mathematical operands simply emerges as a special application: Julia comes with pre-defined catch-all dispatch rules for mathematical operators, invoked when no specific implementation exists for some combination of operand types. These catch-all rules first promote all operands to a common type using user-definable promotion rules, and then invoke a specialized implementation of the operator in question for the resulting values, now of the same type. User-defined types can easily participate in this promotion system by defining methods for conversion to and from other types, and providing a handful of promotion rules defining what types they should promote to when mixed with other types.
Conversion of values to various types is performed by the convert
function. The convert
function generally takes two arguments: the first is a type object while the second is a value to convert to that type; the returned value is the value converted to an instance of given type. The simplest way to understand this function is to see it in action:
julia> x = 12 12 julia> typeof(x) Int64 julia> convert(UInt8, x) 0x0c julia> typeof(ans) UInt8 julia> convert(AbstractFloat, x) 12.0 julia> typeof(ans) Float64 julia> a = Any[1 2 3; 4 5 6] 2×3 Array{Any,2}: 1 2 3 4 5 6 julia> convert(Array{Float64}, a) 2×3 Array{Float64,2}: 1.0 2.0 3.0 4.0 5.0 6.0
Conversion isn’t always possible, in which case a no method error is thrown indicating that convert
doesn’t know how to perform the requested conversion:
julia> convert(AbstractFloat, "foo") ERROR: MethodError: Cannot `convert` an object of type String to an object of type AbstractFloat This may have arisen from a call to the constructor AbstractFloat(...), since type constructors fall back to convert methods. ...
Some languages consider parsing strings as numbers or formatting numbers as strings to be conversions (many dynamic languages will even perform conversion for you automatically), however Julia does not: even though some strings can be parsed as numbers, most strings are not valid representations of numbers, and only a very limited subset of them are. Therefore in Julia the dedicated parse()
function must be used to perform this operation, making it more explicit.
To define a new conversion, simply provide a new method for convert()
. That’s really all there is to it. For example, the method to convert a real number to a boolean is this:
convert(::Type{Bool}, x::Real) = x==0 ? false : x==1 ? true : throw(InexactError())
The type of the first argument of this method is a singleton type, Type{Bool}
, the only instance of which is Bool
. Thus, this method is only invoked when the first argument is the type value Bool
. Notice the syntax used for the first argument: the argument name is omitted prior to the ::
symbol, and only the type is given. This is the syntax in Julia for a function argument whose type is specified but whose value is never used in the function body. In this example, since the type is a singleton, there would never be any reason to use its value within the body. When invoked, the method determines whether a numeric value is true or false as a boolean, by comparing it to one and zero:
julia> convert(Bool, 1) true julia> convert(Bool, 0) false julia> convert(Bool, 1im) ERROR: InexactError() in convert(::Type{Bool}, ::Complex{Int64}) at ./complex.jl:23 ... julia> convert(Bool, 0im) false
The method signatures for conversion methods are often quite a bit more involved than this example, especially for parametric types. The example above is meant to be pedagogical, and is not the actual Julia behaviour. This is the actual implementation in Julia:
convert{T<:Real}(::Type{T}, z::Complex) = (imag(z)==0 ? convert(T,real(z)) : throw(InexactError())) julia> convert(Bool, 1im) ERROR: InexactError() in convert(::Type{Bool}, ::Complex{Int64}) at ./complex.jl:18 ...
To continue our case study of Julia’s Rational
type, here are the conversions declared in rational.jl, right after the declaration of the type and its constructors:
convert{T<:Integer}(::Type{Rational{T}}, x::Rational) = Rational(convert(T,x.num),convert(T,x.den)) convert{T<:Integer}(::Type{Rational{T}}, x::Integer) = Rational(convert(T,x), convert(T,1)) function convert{T<:Integer}(::Type{Rational{T}}, x::AbstractFloat, tol::Real) if isnan(x); return zero(T)//zero(T); end if isinf(x); return sign(x)//zero(T); end y = x a = d = one(T) b = c = zero(T) while true f = convert(T,round(y)); y -= f a, b, c, d = f*a+c, f*b+d, a, b if y == 0 || abs(a/b-x) <= tol return a//b end y = 1/y end end convert{T<:Integer}(rt::Type{Rational{T}}, x::AbstractFloat) = convert(rt,x,eps(x)) convert{T<:AbstractFloat}(::Type{T}, x::Rational) = convert(T,x.num)/convert(T,x.den) convert{T<:Integer}(::Type{T}, x::Rational) = div(convert(T,x.num),convert(T,x.den))
The initial four convert methods provide conversions to rational types. The first method converts one type of rational to another type of rational by converting the numerator and denominator to the appropriate integer type. The second method does the same conversion for integers by taking the denominator to be 1. The third method implements a standard algorithm for approximating a floating-point number by a ratio of integers to within a given tolerance, and the fourth method applies it, using machine epsilon at the given value as the threshold. In general, one should have a//b == convert(Rational{Int64}, a/b)
.
The last two convert methods provide conversions from rational types to floating-point and integer types. To convert to floating point, one simply converts both numerator and denominator to that floating point type and then divides. To convert to integer, one can use the div
operator for truncated integer division (rounded towards zero).
Promotion refers to converting values of mixed types to a single common type. Although it is not strictly necessary, it is generally implied that the common type to which the values are converted can faithfully represent all of the original values. In this sense, the term “promotion” is appropriate since the values are converted to a “greater” type — i.e. one which can represent all of the input values in a single common type. It is important, however, not to confuse this with object-oriented (structural) super-typing, or Julia’s notion of abstract super-types: promotion has nothing to do with the type hierarchy, and everything to do with converting between alternate representations. For instance, although every Int32
value can also be represented as a Float64
value, Int32
is not a subtype of Float64
.
Promotion to a common “greater” type is performed in Julia by the promote
function, which takes any number of arguments, and returns a tuple of the same number of values, converted to a common type, or throws an exception if promotion is not possible. The most common use case for promotion is to convert numeric arguments to a common type:
julia> promote(1, 2.5) (1.0,2.5) julia> promote(1, 2.5, 3) (1.0,2.5,3.0) julia> promote(2, 3//4) (2//1,3//4) julia> promote(1, 2.5, 3, 3//4) (1.0,2.5,3.0,0.75) julia> promote(1.5, im) (1.5 + 0.0im,0.0 + 1.0im) julia> promote(1 + 2im, 3//4) (1//1 + 2//1*im,3//4 + 0//1*im)
Floating-point values are promoted to the largest of the floating-point argument types. Integer values are promoted to the larger of either the native machine word size or the largest integer argument type. Mixtures of integers and floating-point values are promoted to a floating-point type big enough to hold all the values. Integers mixed with rationals are promoted to rationals. Rationals mixed with floats are promoted to floats. Complex values mixed with real values are promoted to the appropriate kind of complex value.
That is really all there is to using promotions. The rest is just a matter of clever application, the most typical “clever” application being the definition of catch-all methods for numeric operations like the arithmetic operators +
, -
, *
and /
. Here are some of the catch-all method definitions given in promotion.jl:
+(x::Number, y::Number) = +(promote(x,y)...) -(x::Number, y::Number) = -(promote(x,y)...) *(x::Number, y::Number) = *(promote(x,y)...) /(x::Number, y::Number) = /(promote(x,y)...)
These method definitions say that in the absence of more specific rules for adding, subtracting, multiplying and dividing pairs of numeric values, promote the values to a common type and then try again. That’s all there is to it: nowhere else does one ever need to worry about promotion to a common numeric type for arithmetic operations — it just happens automatically. There are definitions of catch-all promotion methods for a number of other arithmetic and mathematical functions in promotion.jl, but beyond that, there are hardly any calls to promote
required in the Julia standard library. The most common usages of promote
occur in outer constructors methods, provided for convenience, to allow constructor calls with mixed types to delegate to an inner type with fields promoted to an appropriate common type. For example, recall that rational.jl provides the following outer constructor method:
Rational(n::Integer, d::Integer) = Rational(promote(n,d)...)
This allows calls like the following to work:
julia> Rational(Int8(15),Int32(-5)) -3//1 julia> typeof(ans) Rational{Int32}
For most user-defined types, it is better practice to require programmers to supply the expected types to constructor functions explicitly, but sometimes, especially for numeric problems, it can be convenient to do promotion automatically.
Although one could, in principle, define methods for the promote
function directly, this would require many redundant definitions for all possible permutations of argument types. Instead, the behavior of promote
is defined in terms of an auxiliary function called promote_rule
, which one can provide methods for. The promote_rule
function takes a pair of type objects and returns another type object, such that instances of the argument types will be promoted to the returned type. Thus, by defining the rule:
promote_rule(::Type{Float64}, ::Type{Float32} ) = Float64
one declares that when 64-bit and 32-bit floating-point values are promoted together, they should be promoted to 64-bit floating-point. The promotion type does not need to be one of the argument types, however; the following promotion rules both occur in Julia’s standard library:
promote_rule(::Type{UInt8}, ::Type{Int8}) = Int promote_rule(::Type{BigInt}, ::Type{Int8}) = BigInt
In the latter case, the result type is BigInt
since BigInt
is the only type large enough to hold integers for arbitrary-precision integer arithmetic. Also note that one does not need to define both promote_rule(::Type{A}, ::Type{B})
and promote_rule(::Type{B}, ::Type{A})
— the symmetry is implied by the way promote_rule
is used in the promotion process.
The promote_rule
function is used as a building block to define a second function called promote_type
, which, given any number of type objects, returns the common type to which those values, as arguments to promote
should be promoted. Thus, if one wants to know, in absence of actual values, what type a collection of values of certain types would promote to, one can use promote_type
:
julia> promote_type(Int8, UInt16) Int64
Internally, promote_type
is used inside of promote
to determine what type argument values should be converted to for promotion. It can, however, be useful in its own right. The curious reader can read the code in promotion.jl, which defines the complete promotion mechanism in about 35 lines.
Finally, we finish off our ongoing case study of Julia’s rational number type, which makes relatively sophisticated use of the promotion mechanism with the following promotion rules:
promote_rule{T<:Integer,S<:Integer}(::Type{Rational{T}}, ::Type{S}) = Rational{promote_type(T,S)} promote_rule{T<:Integer,S<:Integer}(::Type{Rational{T}}, ::Type{Rational{S}}) = Rational{promote_type(T,S)} promote_rule{T<:Integer,S<:AbstractFloat}(::Type{Rational{T}}, ::Type{S}) = promote_type(T,S)
The first rule says that promoting a rational number with any other integer type promotes to a rational type whose numerator/denominator type is the result of promotion of its numerator/denominator type with the other integer type. The second rule applies the same logic to two different types of rational numbers, resulting in a rational of the promotion of their respective numerator/denominator types. The third and final rule dictates that promoting a rational with a float results in the same type as promoting the numerator/denominator type with the float.
This small handful of promotion rules, together with the conversion methods discussed above, are sufficient to make rational numbers interoperate completely naturally with all of Julia’s other numeric types — integers, floating-point numbers, and complex numbers. By providing appropriate conversion methods and promotion rules in the same manner, any user-defined numeric type can interoperate just as naturally with Julia’s predefined numerics.
© 2009–2016 Jeff Bezanson, Stefan Karpinski, Viral B. Shah, and other contributors
Licensed under the MIT License.
http://docs.julialang.org/en/release-0.5/manual/conversion-and-promotion/