This document explains the low-level inner workings of
diagrams-core. Casual users of diagrams should not need to
read this (although a quick skim may well turn up something
interesting). It is intended more for developers and power users who
want to learn how diagrams actually works under the hood; there is
quite a lot that goes on behind the scenes to enable the powerful
tools provided in diagrams-lib.
Chris Mears has written an article giving a quick walkthrough of some of the internals which is useful for getting started.
The remainder of this document is organized around the modules in
diagrams-core. At some level, there is no substitute for just
diving in and reading the source code (see the diagrams-core
repository), which is generally well-commented, but the hope is that
this document can serve to orient you and supply useful commentary.
This module simply re-exports many things from the other modules for convenience.
Diagrams.Core.V contains the definition of the V type family,
which maps from types to their associated vector space.
(See the relevant section from the user manual) along with some
basic instances.
The linear package defines a Point type in Linear.Affine
along with some functions for working with points.
The Diagrams.Core.Points module simply re-exports a few things
from linear, defines an instance of V and N for Point,
and adds a few utility functions for points.
Diagrams.Core.Names defines the infrastructure for names which
can be used to identify subdiagrams.
AName, representing atomic names, is an existential wrapper,
allowing (almost) any type to be used for names, as the user finds
convenient. Strings may be used of course, but also numbers,
characters, even user-defined types. The only restriction is that the
wrapped type must be an instance of the following three classes:
Typeable (so values can be pulled back out of the wrappers in a
type-safe way),
Show (so names can be displayed, for debugging purposes), and
Ord (in order to be able to create a Map from names to
subdiagrams).
Equality on atomic names works as expected: two names are equal if their types match and their values are equal.
The Ord instance for atomic names works by first ordering names
according to (a String representation of) their type, and then by
value for equal types (using the required Ord instance).
A qualified name (Name) is a list of atomic names. The IsName
class covers things which can be used as a name, including many
standard base types as well as ANames and Names. Most user-facing
functions which take a name as an argument actually take any type with
an IsName constraint, so the user can just pass in a String or an
Int or whatever they want.
The motivation for having names consist of lists of atomic names is
that it is not always convenient or even feasible to have globally
unique names (especially when multiple modules by different authors
are involved). In such a situation it is possible to qualify all
the names in a particular diagram by some prefix. This operation
governed by the Qualifiable class, containing the function (|>) ::
IsName a => a -> q -> q for performing qualification.
This module defines the HasOrigin class (containing the
moveOriginTo method) as well as related functions like
moveOriginBy, moveTo, and place. It also defines instances of
HasOrigin for a number of types, including Points, tuples, lists,
sets, and maps.
See the section of the type class reference on HasOrigin for more information.
This module defines a type of generic affine transformations parameterized over any vector space, along with a large number of methods for working with transformations.
First, the (:-:) type consists of a pair of functions, which are
assumed to be linear and inverse to each other.
A Transformation type is then defined to contain three components:
a linear map and its inverse (stored using (:-:))
the transpose of the linear map, with its inverse (again stored using (:-:))
a vector, representing a translation
The point is that we need transposes and inverses when transforming
things like Envelopes and Traces. While it would be possible in
theory to simply store a transformation as a matrix and compute its
transpose or inverse whenever required, this would be computationally
wasteful (especially computing inverses). Instead, we simply package
up a transformation along with its inverse, transpose, and inverse
transpose (which we can think of as a little 2x2 matrix of functions).
Such a representation is closed under composition, and we can compute
its inverse or transpose by just flipping the matrix along the
appropriate axis.
Along with the definition of the Transformation type itself, this
module exports many functions for generically creating, transforming,
querying, and applying Transformation values. For example, in
addition to straightforward things like composing and applying
transformations, this is where you can find code to convert a
Transformation to a matrix representation or to compute its
determinant. (On the other hand, converting a matrix to a
Transformation is only implemented specifically for 2 or 3
dimensions, and can be found in the diagrams-lib package, in
Diagrams.Transform.Matrix.)
This module also defines the important Transformable class of things
to which Transformations can be applied, along with many generic
instances.
Finally, the module defines a few specific transformations which are
polymorphic over the vector space, namely, translation and scaling.
Other specific transformations (e.g. scaleX and so on) are defined
in diagrams-lib.
This module defines the Envelope type; see the user manual section
on envelopes for a general overview of what envelopes are and how
to use them.
For an explanation of the specific way that Envelope is defined, see
Brent Yorgey's paper on diagrams and monoids.
The real meat of this module consists of the definitions of
HasOrigin and Transformable instances for the Envelope type.
The fact that packaging transformations together with their transpose
and inverse makes it possible to correctly compute the affine
transformation of an envelope is one of the key insights making the
diagrams framework possible. The source code has extensive comments
explaining the instances; consult those if you want to understand
how they actually work.
Finally, this module defines the Enveloped class for things with
Envelopes, a number of functions like envelopeV, envelopePMay,
and so on for querying envelopes, and size-related functions like
diameter, extent, and size that are defined in terms of
envelopes.
This module defines the Juxtaposable class, the default
implementation juxtaposeDefault for instances of Enveloped and
HasOrigin, and generic instances for Envelope, pairs, lists, maps,
sets, and functions.
See the type class reference section on Juxtaposable for more information.
This module defines the Measured type along with a number of utility
functions and instances for working with it. See the user manual
section on measurement units.
Measured values are implemented as functions from a triple of
scaling factors to a final value: the local scaling factor, global
scaling factor, and normalized scaling factor. XXX write about how
these are computed
This module implements the trace which is associated with every diagram. A trace is essentially an "embedded raytracer" which can compute an intersection with a diagram in any direction from any given base point. Note that a trace needs to be able to answer a trace query from any given base point, not just from some chosen particular base point (e.g. the origin), since we need to be able to apply affine transformations, including translations.
Often when one thinks about raytracing the basic idea is that you
follow a ray and return the first intersection that occurs.
However, to allow for also computing the last intersection and other
generalizations, the base framework in this module actually computes a
sorted list of all the intersection points. Hence this module
defines a small abstraction for sorted lists, as well as the Trace
abstraction itself. A number of functions for querying Trace values
are defined here, as well as the Traced class for things which have
a Trace.
A Query is a function that associates a value of some (monoidal)
type to each point in a diagram; see the user manual section on
queries. There is not much in this module besides a great many
type class instances for the Query type.
This module implements styles, which are collections of attributes (such as line color, fill color, opacity, ...) that can be applied to diagrams. Diagrams takes a dynamically typed approach to attributes and styles. This is in contrast to the approach with backends and primitives, where the type of a diagram tells you what backend it is to be rendered with—or, if it is polymorphic in the backend, there are type class constraints that say what primitives the backend must be able to render. But the type of a diagram never says anything about what attributes a backend must support; indeed, by looking only at the type of a diagram it is impossible to tell what types of attributes it contains. In general, backends pick out the attributes they can handle and simply ignore any others.
Attributes are the primitive values out of which styles are built.
Almost any type can be used as an attribute, with only a few
restrictions: attributes must be Typeable, to support the use of
dynamic typing, and a Semigroup, so there is some sensible notion of
combining multiple attributes of the same type (which is used to
combine attributes applied within the same scope; as we will see, for
many standard attributes the semigroup is simply the one which keeps
one attribute and discards the other). AttributeClass is defined as
a synonym for the combination of Typeable and Semigroup.
The Attribute type is then defined as an existential wrapper around
AttributeClass types. In a simpler world Attribute would be
defined like this:
> data Attribute where
> Attribute :: AttributeClass a => a -> AttributeHistorically, it did indeed start life defined this way. However, as you can see if you look at the source, by now the actual definition is more complicated:
> data Attribute (v :: * -> *) n :: * where
> Attribute :: AttributeClass a => a -> Attribute v n
> MAttribute :: AttributeClass a => Measured n a -> Attribute v n
> TAttribute :: (AttributeClass a, Transformable a, V a ~ v, N a ~ n) => a -> Attribute v nThis looks like the simpler definition if you ignore the type
parameters and consider only the Attribute constructor. So let's
consider each of the other constructors.
MAttribute is for attributes that are Measured, i.e. whose
values depend on the size of the final diagram and/or the requested
output size; the primary examples are line width and font size.
Recall that a Measured n a is actually a function that can produce
a value of type a once it is provided some measurement factors of
type n. The unmeasureAttribute function is provided to turn
MAttribute constructors into Attribute constructors; this is
typically used when preparing a diagram for rendering.
TAttribute is for attributes that are Transformable, i.e.
which are affected by transformations applied to the objects to
which they are attached. The primary examples are line and fill
texture (e.g. gradients), and clipping paths. (Note that
MAttributes can actually be affected by transformations too, in
the case of Local units.)
The Attribute type has instances of Semigroup (combine attributes
of the same type, otherwise take the rightmost) and Transformable
(ignore Attribute constructors and do the appropriate thing for the
other constructors). There are also various lenses/prisms for
accessing them.
Note that one does not typically construct an Attribute value directly
using the constructors; instead, the functions applyAttr,
applyMAttr, and applyTAttr are provided for applying an attribute
directly to any instance of HasStyle.
A Style is just a dynamically-typed, heterogeneous collection of
attributes. The attributes are actually stored as values in a hash
table, keyed by their type (specifically, a TypeRep), so at most one
attribute of any given type can be stored in a Style.
Style is an instance of Semigroup, with union as a combining
operation. If the two styles both have an attribute of a given type,
those attributes are combined according to the Semigroup instance
for that attribute type.
There are a number of other functions for creating Style values and
extracting attributes from them, which are mostly self-explanatory.
One function worth mentioning is unmeasureAttrs, which maps over a
Style and changes all MAttributes into Attributes (based on the
provided scaling factors). This is typically done as a final step
before rendering.
Finally, the HasStyle class governs types which "have a Style",
specifically, types to which a Style can be applied. Style itself
has an instance of HasStyle, corresponding to the semigroup
operation on Style. This module also defines a number of other
instances for applying styles to entire data structures such as lists,
tuples, functions, maps, and sets.
This is an unfortunately large module which contains definitions and utility functions for many of the core data structures of diagrams. In principle, it would be nice to break it up into smaller pieces, but in fact a lot of things in this module end up cyclically depending on one another, so easier said than done.
The central type in diagrams is the definition of QDiagram:
> newtype QDiagram b v n m
> = QD (D.DUALTree (DownAnnots v n) (UpAnnots b v n m) Annotation (QDiaLeaf b v n m))DUALTree is defined in the dual-tree package. A value of
type DUALTree d u a l consists of an n-ary (rose) tree, with:
Values of type l at the leaves.
Monoidal annotations of type u which "travel up" the tree. Each
leaf of type l has a corresponding value of type u, and the u
values are combined as one travels up the tree, so that the root
would contain the mconcat of the u values in all the leaves. We
think of each tree (and subtree) as intrinsically possessing a u
value, which is often some kind of "summary" or "measurement".
There are also values of another monoid, of type d, which can be
applied to a tree at the root and accumulate as one travels down
any path from the root to a leaf. These d values may act on the u
values, that is, there is a function d -> u -> u satisfying
certain coherence properties with respect to the monoid structures
of d and u (though for a particular d and u the action may
be trivial, i.e. const id).
Finally, there are values of type a which can be stored at
internal nodes. They are simply "along for the ride": they do not
have associated u values nor are they affected by d values. a
values will never be moved (on the other hand, it is permissible to
push d values up or down the tree in a way that preserves all
monoid compositions).
The QDiagram type specifically instantiates these types as follows:
l values at leaves are QDiaLeaf values. One might think these
consist simply of primitives, but actually they are a bit more
complicated, because they also handle the case of "delayed" diagrams
that need to know something about their context before they can be
generated; this is explained in more detail below.
The inert internal a values are of type Annotation. Currently
these consist solely of information about hrefs (for backends that
support hyperlinks) and opacity grouping.
The upwards-traveling u values are of type UpAnnots, explained
in more detail below, which record various summary information about
a diagram (envelope, trace, named subdiagrams, etc.).
The downwards-traveling d values are of type DownAnnots,
explained in more detail below, which record information imposed on
a diagram (affine transformations, attributes, etc.).
The QDiaLeaf type has two cases:
> data QDiaLeaf b v n m
> = PrimLeaf (Prim b v n)
> | DelayedLeaf (DownAnnots v n -> n -> n -> QDiagram b v n m)The first case, PrimLeaf, is simple enough: it contains a Prim
(also defined in this module), which is simply an existential wrapper
around a Transformable, Typeable, Renderable thing.
The second case is trickier. It represents a diagram which needs to
know its global context, represented by the accumulated DownAnnots
values along the path from the root to this delayed leaf, together
with the normal-to-output scaling factor and the global-to-output
scaling factor. Note that we cannot possibly know this information
until just before the diagram is to be rendered, at which point it
becomes fixed and we know we are not going to insert this diagram into
a yet bigger diagram, and we know the requested output size and hence
can compute the scaling factors. So just before rendering, delayed
leaves are recursively expanded into QDiagrams. For an example of
DelayedLeaf in action, see Diagrams.TwoD.Arrow in
diagrams-lib, where it is needed since arrowhead sizes can
depend on this global context.
The Annotation type has two constructors, one for hyperlinks
(Href) and one for opacity groups. In both of these cases, it's
important that these annotations stay exactly where the user places
them in the tree, and they are unaffected by transformations and the
like.
UpAnnots is a heterogeneous list of monoidal values (with an
elementwise monoidal operation; see Data.Monoid.MList in the
monoid-extras package) which serves as the "upwards-traveling"
monoid in a QDiagram. That is, every primitive diagram at a leaf
has an intrinsic associated UpAnnots value, and these get combined
as one moves up the tree. An UpAnnots value consists of the
following components:
An Envelope (in a Deletable wrapper so we can implement
setEnvelope monoidally; see Data.Monoid.Deletable in
monoid-extras),
a Trace (similarly Deletable),
a SubMap (also Deletable) which maps names to subdiagrams (more
on these later), and
a Query, which associates a monoidal value of some type m to
every point in the diagram.
DownAnnots is