EltypeExtensions.jl is a mini toolbox for eltype-related conversions. The motivation of this package comes from manipulating (nested) arrays with different eltypes. However if you have any reasonable idea that works on other instances, feel free to write an issue/pull request.
We note that this package has some overlap with TypeUtils.jl and Unitless.jl.
elconvert(T, x)
works like convert(T, x)
, except that T
refers to the eltype of the result. This can be useful for generic codes.
It should be always true that elconvert(T, x) isa _to_eltype(T, typeof(x))
. However, since elconvert
and _to_eltype
use different routines, it's possible that the equality doesn't hold for some types. Please submit an issue or PR if that happens.
If typeof(x)
is not in Base or stdlib, the package who owns the type should implement corresponding _to_eltype
or elconvert
. elconvert
has fallbacks, in which case it could be unnecessary:
- For a subtype of
AbstractArray
,elconvert
calls the constructorAbstractArray{T}
and_to_eltype
returnsArray
. - For a subtype of
AbstractUnitRange
,elconvert
calls the constructorAbstractUnitRange{T}
. - For a subtype of
AbstractRange
,elconvert
uses broadcast throughmap
. - For a
Tuple
,elconvert
uses dot broadcast. - For other types,
elconvert
callsconvert
and_to_eltype
.
However, _to_eltype
must be implemented for each type to support baseconvert
and precisionconvert
. The following types from Base and stdlib are explicitly supported by _to_eltype
:
AbstractArray, AbstractDict, AbstractSet, Adjoint, Bidiagonal, BitArray,
CartesianIndices, Diagonal, Dict, Hermitian, Set, StepRangeLen, Symmetric,
SymTridiagonal, Transpose, TwicePrecision, UnitRange
The basetype
is used for nested collections, where eltype
is repeatedly applied until the bottom. precisiontype
has a similar idea, but goes deeper when possible. precisiontype
is used to manipulate the accuracy of (nested) collections.
sometype(T)
gets thesometype
of typeT
.sometype(x) = sometype(typeof(x))
is also provided for convenience._to_sometype(T,S)
converts the typeS
to have thesometype
ofT
.someconvert(T,A)
convertsA
to have thesometype
ofT
.
where some
can be el
, base
and precision
.
precisionconvert
accepts an optional third argument prec
.
- When
T
has static precision,prec
has no effect. - When
T
has dynamic precision,prec
specifies the precision of conversion. Whenprec
is not provided, the precision is decided by the external setup fromT
. The difference is significant whenprecisionconvert
is called by another function. See the document for an example. - When
T
is an integer, the conversion will dig intoRational
as well. In contrast, sinceRational
as a whole is more "precise" than an integer,precisiontype
doesn't unwrapRational
.