\chapter[chap:description]{Hardware description}
This chapter will provide an overview of the hardware description language
that was created and the issues that have arisen in the process. It will
focus on the issues of the language, not the implementation.
When translating Haskell to hardware, we need to make choices in what kind
of hardware to generate for what Haskell constructs. When faced with
choices, we've tried to stick with the most obvious choice wherever
possible. In a lot of cases, when you look at a hardware description it is
comletely clear what hardware is described. We want our translator to
generate exactly that hardware whenever possible, to minimize the amount of
surprise for people working with it.
In this chapter we try to describe how we interpret a Haskell program from a
hardware perspective. We provide a description of each Haskell language
element that needs translation, to provide a clear picture of what is
supported and how.
\section{Function application}
The basic syntactic element of a functional program are functions and
function application. These have a single obvious \small{VHDL} translation: Each
function becomes a hardware component, where each argument is an input port
and the result value is the output port.
Each function application in turn becomes component instantiation. Here, the
result of each argument expression is assigned to a signal, which is mapped
to the corresponding input port. The output port of the function is also
mapped to a signal, which is used as the result of the application.
\in{Example}[ex:And3] shows a simple program using only function
application and the corresponding architecture.
\startbuffer[And3]
-- | A simple function that returns
-- the and of three bits
and3 :: Bit -> Bit -> Bit -> Bit
and3 a b c = and (and a b) c
\stopbuffer
\startuseMPgraphic{And3}
save a, b, c, anda, andb, out;
% I/O ports
newCircle.a(btex $a$ etex) "framed(false)";
newCircle.b(btex $b$ etex) "framed(false)";
newCircle.c(btex $c$ etex) "framed(false)";
newCircle.out(btex $out$ etex) "framed(false)";
% Components
newCircle.anda(btex $and$ etex);
newCircle.andb(btex $and$ etex);
a.c = origin;
b.c = a.c + (0cm, 1cm);
c.c = b.c + (0cm, 1cm);
anda.c = midpoint(a.c, b.c) + (2cm, 0cm);
andb.c = midpoint(b.c, c.c) + (4cm, 0cm);
out.c = andb.c + (2cm, 0cm);
% Draw objects and lines
drawObj(a, b, c, anda, andb, out);
ncarc(a)(anda) "arcangle(-10)";
ncarc(b)(anda);
ncarc(anda)(andb);
ncarc(c)(andb);
ncline(andb)(out);
\stopuseMPgraphic
\placeexample[here][ex:And3]{Simple three port and.}
\startcombination[2*1]
{\typebufferhs{And3}}{Haskell description using function applications.}
{\boxedgraphic{And3}}{The architecture described by the Haskell description.}
\stopcombination
TODO: Define top level function and subfunctions/circuits.
\section{Choice}
Since describing components and connections allows us to describe a lot of
hardware designs already, there is an obvious thing missing: Choice. We
need some way to be able to choose between values based on another value.
In Haskell, choice is achieved by case expressions, if expressions or
pattern matching (all of which can be reduced to case expressions).
An obvious way to add choice to our language without having to recognize
any of Haskell's syntax, would be to add a primivite \quote{\hs{if}}
function. This function would take three arguments: The condition, the
value to return when the condition is true and the value to return when
the condition is false.
This \hs{if} function would then essentially describe a multiplexer and
allows us to describe any architecture that uses multiplexers. (TODO: Are
there other mechanisms of choice in hardware?)
However, to be able to describe our hardware in a more convenient way, we
also need to translate Haskell's choice mechanisms. The easiest of these
are of course case expressions (and if expressions, which can be very
directly translated to case expressions). A case expression can in turn
simply be translated to a conditional assignment, where the conditions use
equality comparisons against the constructors in the case expressions.
In \in{example}[ex:CaseInv] a simple case statement is shown,
scrutinizing a boolean value. If we would translate a Boolean to a bit
value, we could of course remove the comparator and directly feed 'in'
into the multiplex (or even use an inverter instead of a multiplexer).
However, we will try to make a general translation, which works for all
possible case statements. Optimizations such as these are left for the
\VHDL synthesizer, which handles them very well.
\startbuffer[CaseInv]
inv :: Bool -> Bool
inv in = case in of
True -> False
False -> True
\stopbuffer
\startuseMPgraphic{CaseInv}
save in, truecmp, falseout, trueout, out, cmp, mux;
% I/O ports
newCircle.in(btex $in$ etex) "framed(false)";
newCircle.out(btex $out$ etex) "framed(false)";
% Constants
newBox.truecmp(btex $True$ etex) "framed(false)";
newBox.trueout(btex $True$ etex) "framed(false)";
newBox.falseout(btex $False$ etex) "framed(false)";
% Components
newCircle.cmp(btex $==$ etex);
newMux.mux;
in.c = origin;
cmp.c = in.c + (3cm, 0cm);
truecmp.c = cmp.c + (-1cm, 1cm);
mux.sel = cmp.e + (1cm, -1cm);
falseout.c = mux.inpa - (2cm, 0cm);
trueout.c = mux.inpb - (2cm, 0cm);
out.c = mux.out + (2cm, 0cm);
% Draw objects and lines
drawObj(in, out, truecmp, trueout, falseout, cmp, mux);
ncline(in)(cmp);
ncarc(truecmp)(cmp);
nccurve(cmp.e)(mux.sel) "angleA(0)", "angleB(-90)";
ncline(falseout)(mux) "posB(inpa)";
ncline(trueout)(mux) "posB(inpb)";
ncline(mux)(out) "posA(out)";
\stopuseMPgraphic
\placeexample[here][ex:CaseInv]{Simple inverter.}
\startcombination[2*1]
{\typebufferhs{CaseInv}}{Haskell description using a Case expression.}
{\boxedgraphic{CaseInv}}{The architecture described by the Haskell description.}
\stopcombination
Slightly more complex (but very powerful) forms of choice are pattern
matches. A function can be defined in multiple clauses, where each clause
specifies a pattern. When the arguments match the pattern, the
corresponding clause will be used.
\startbuffer[PatternInv]
inv :: Bool -> Bool
inv True = False
inv False = True
\stopbuffer
\placeexample[here][ex:PatternInv]{Simple inverter using pattern matching.}
{\typebufferhs{CaseInv}}
The architecture described in \in{example}[ex:PatternInv] is of course the
same one as the one in \in{example}[ex:CaseInv]. The general interpration
of pattern matching is also similar to that of case expressions: Generate
hardware for each of the clause (like each of the clauses of a case
expression) and connect them to the function output through (a number of
nested) multiplexers. These multiplexers are driven by comparators and
other logic, that check each pattern in turn.
\section{Types}
Apart from giving structure to our hardware, we'll also need to provide
some way to translate the values used to hardware equivalents. In
particular, this means having to come up with a hardware equivalent for
every \emph{type} used in our program.
Since most functional languages have a lot of standard types that are
hard to translate (integers without a fixed size, lists without a static
length, etc.), we will start out by defining a number of \quote{builtin}
types ourselves. These types are builtin in the sense that our compiler
will have a fixed VHDL type for these. User defined types, on the other
hand, will have their hardware type derived directly from their Haskell
declaration automatically, according to the rules we sketch here.
\subsection{Builtin types}
\startdesc{Bit}
This is the most basic type available. It is mapped directly onto
the \type{std_logic} \small{VHDL} type. Mapping this to the
\type{bit} type might make more sense (since the Haskell version
only has two values), but using \type{std_logic} is more standard
(and allowed for some experimentation with don't care values)
\stopdesc
\startdesc{Bool}
This is a builtin Haskell type, but is translated exactly like the
Bit type. Supporting the Bool type is particularly useful to support
\hs{if ... then ... else ...} expressions, which always have a
\hs{Bool} value for the condition.
A \hs{Bool} is translated to a \type{std_logic}, just like \hs{Bit}.
\stopdesc
\startdesc{SizedWord, SizedInt}
These are types to represent integers. \hs{SizedWord} is unsigned,
while \hs{SizedInt} is signed. These types are parameterized by a
length type, so you can define an unsigned word of 32 bits wide as
follows:
\starthaskell
type Word32 = SizedWord D32
\stophaskell
Here, \hs{D32} is the \emph{type level representation} of the number
32. TODO: Introduce dependent types somewhere.
These types are translated to the \small{VHDL} \type{unsigned} and
\type{signed} respectively.
\stopdesc
\startdesc{RangedWord}
This is another type to describe integers, but unlike the previous
two it has no specific width, but an upper bound. This means that
its range is not limited to powers of two, but can be any number.
A \hs{RangedWord} only has an upper bound, its lower bound is
implicitly zero. There is a lot of added implementation complexity
when adding a lower bound and having just an upper bound was enough
for the primary purpose of this type: Typesafely indexing vectors.
This type is translated to the \type{unsigned} \small{VHDL} type.
\stopdesc
\startdesc{TFVec}
This is a vector type, with a fixed length. It has two type
parameters: Its length and the type of the elements contained in it.
A fixed size vector is translated to a \small{VHDL} array type.
The \hs{TF} in its name is a reference to the implementation used in
the prototype, which uses \emph{type families}.
\stopdesc
TODO: Reference Christiaan / describe dependent typing
\subsection{User-defined types}
There are three ways to define new types in Haskell: Algebraic
datatypes with the \hs{data} keyword, type synonyms with the \hs{type}
keyword and type renamings with the \hs{newtype} keyword. This
expclitely excludes more advanced type creation from \GHC extensions
such as type families, existential typing, \small{GADT}s, etc.
The first of these actually introduces a new type, for which we provide
the \VHDL translation below. The latter two only define new names for
existing types (where synonyms are completely interchangeable and
renamings need explicit conversion). Therefore, these don't need any
particular \VHDL translation, a synonym or renamed type will just use
the same representation as the equivalent type.
For algebraic types, we can make the following distinction:
\startdesc{Product types}
A product type is an algebraic datatype with a single constructor with
two or more fields. This is essentially a way to pack a few values
together in a record-like structure. In fact, the builtin tuple types
are just product types (and are thus supported in exactly the same
way).
The "product" in its name refers to the collection of values belonging
to this type. The collection for a product type is the cartesic
product of the collections for the types of its fields.
These types are translated to \VHDL record types, with one field for
every field in the constructor. This translation applies to all single
constructor algebraic datatypes, including those with no fields (unit
types) and just one field (which are technically not a product).
\stopdesc
\startdesc{Enumerated types}
An enumerated type is an algebraic datatype with multiple constructors, but
none of them have fields. This is essentially a way to get an
enum-like type containing alternatives.
Note that Haskell's \hs{Bool} type is also defined as an
enumeration type, but we have a fixed translation for that.
These types are translated to \VHDL enumerations, with one value for
each constructor. This allows references to these constructors to be
translated to the corresponding enumeration value.
\stopdesc
\startdesc{Sum types}
A sum type is an algebraic datatype with multiple constructors, where
the constructors have one or more fields. Technically, a type with
more than one field per constructor is a sum of products type, but
for our purposes this distinction does not really make a difference,
so we'll leave it out.
Sum types are currently not supported by the prototype, since its
translation is not so obvious as before. The can easily be emulated,
as we will see from an example:
\starthaskell
data Sum = A Bit Word | B Word
\stophaskell
An obvious way to translate this would be to create an enumeration to
distinguish the constructors and then create a big record that
contains all the fields of all the constructors. This is the same
translation that would result from the following enumeration and
product type (using a tuple for clarity):
\starthaskell
data SumC = A | B
type Sum = (SumC, Bit, Word, Word)
\stophaskell
Here, the \hs{SumC} type effectively signals which of the fields of
the \hs{Sum} type are valid, all the other ones have no useful value.
An obvious problem with this naive approach is the space usage: The
example above generates a fairly big \VHDL type. However, we can be
sure that the two \hs{Word}s in the \hs{Sum} type will never be valid
at the same time, so this is a waste of space.
Obviously, we could do some duplication detection here to reuse a
particular field for another constructor, but this would only
partially solve the problem. If I would, for example, have an array of
8 bits and a 8 bit unsiged word, these are different types and could
not be shared. However, in the final hardware, both of these types
would simply be 8 bit connections, so we have a 100\% size increase by
not sharing these!
\stopdesc
Another interesting case is that of recursive types. In Haskell, an
algebraic datatype can be recursive: Any of its field types can be (or
contain) the type being defined. The most well-known recursive type is
probably the list type, which is defined is:
\starthaskell
data List a = Empty | Cons a (List a)
\stophaskell
Where \hs{Empty} is usually written as \hs{[]} and \hs{Cons} as \hs{:}.
This immediately shows the problem with recursive types: What hardware
type to allocate here? There is no way to statically determine how long
the list will be
In general, we can say we can never properly translated recursive types:
All recursive types have a potentially infinite value (even though in
practice they will have a bounded value, there is no way for the
compiler to determine an upper bound on its size.
\subsection{Partial application}
It should be obvious that we cannot generate hardware signals for all
expressions we can express in Haskell. The most obvious criterium for this
is the type of an expression. We will see more of this below, but for now it
should be obvious that any expression of a function type cannot be
represented as a signal or i/o port to a component.
From this, we can see that the above translation rules do not apply to a
partial application. \in{Example}[ex:Quadruple] shows an example use of
partial application and the corresponding architecture.
\startbuffer[Quadruple]
-- | Multiply the input word by four.
quadruple :: Word -> Word
quadruple n = mul (mul n)
where
mul = (*) 2
\stopbuffer
\startuseMPgraphic{Quadruple}
save in, two, mula, mulb, out;
% I/O ports
newCircle.in(btex $n$ etex) "framed(false)";
newCircle.two(btex $2$ etex) "framed(false)";
newCircle.out(btex $out$ etex) "framed(false)";
% Components
newCircle.mula(btex $\times$ etex);
newCircle.mulb(btex $\times$ etex);
two.c = origin;
in.c = two.c + (0cm, 1cm);
mula.c = in.c + (2cm, 0cm);
mulb.c = mula.c + (2cm, 0cm);
out.c = mulb.c + (2cm, 0cm);
% Draw objects and lines
drawObj(in, two, mula, mulb, out);
nccurve(two)(mula) "angleA(0)", "angleB(45)";
nccurve(two)(mulb) "angleA(0)", "angleB(45)";
ncline(in)(mula);
ncline(mula)(mulb);
ncline(mulb)(out);
\stopuseMPgraphic
\placeexample[here][ex:Quadruple]{Simple three port and.}
\startcombination[2*1]
{\typebufferhs{Quadruple}}{Haskell description using function applications.}
{\boxedgraphic{Quadruple}}{The architecture described by the Haskell description.}
\stopcombination
Here, the definition of mul is a partial function application: It applies
\hs{2 :: Word} to the function \hs{(*) :: Word -> Word -> Word} resulting in
the expression \hs{(*) 2 :: Word -> Word}. Since this resulting expression
is again a function, we can't generate hardware for it directly. This is
because the hardware to generate for \hs{mul} depends completely on where
and how it is used. In this example, it is even used twice!
However, it is clear that the above hardware description actually describes
valid hardware. In general, we can see that any partial applied function
must eventually become completely applied, at which point we can generate
hardware for it using the rules for function application above. It might
mean that a partial application is passed around quite a bit (even beyond
function boundaries), but eventually, the partial application will become
completely applied.
\section{Costless specialization}
Each (complete) function application in our description generates a
component instantiation, or a specific piece of hardware in the final
design. It is interesting to note that each application of a function
generates a \emph{separate} piece of hardware. In the final design, none
of the hardware is shared between applications, even when the applied
function is the same.
This is distinctly different from normal program compilation: Two separate
calls to the same function share the same machine code. Having more
machine code has implications for speed (due to less efficient caching)
and memory usage. For normal compilation, it is therefore important to
keep the amount of functions limited and maximize the code sharing.
When generating hardware, this is hardly an issue. Having more \quote{code
sharing} does reduce the amount of \small{VHDL} output (Since different
component instantiations still share the same component), but after
synthesis, the amount of hardware generated is not affected.
In particular, if we would duplicate all functions so that there is a
duplicate for every application in the program (\eg, each function is then
only applied exactly once), there would be no increase in hardware size
whatsoever.
TODO: Perhaps these next two sections are a bit too
implementation-oriented?
\subsection{Specialization}
Because of this, a common optimization technique called
\emph{specialization} is as good as free for hardware generation. Given
some function that has a \emph{domain} of $D$ (\eg, the set of all
possible arguments that could be applied), we create a specialized
function with exactly the same behaviour, but with an domain of $D'
\subset D$. This subset can be derived in all sort of ways, but commonly
this is done by limiting a polymorphic argument to a single type (\eg,
removing polymorphism) or by limiting an argument to just a single value
(\eg, cross-function constant propagation, effectively removing the
argument).
Since we limit the argument domain of the specialized function, its
definition can often be optimized further (since now more types or even
values of arguments are already know). By replacing any application of
the function that falls within the reduced domain by an application of
the specialized version, the code gets faster (but the code also gets
bigger, since we now have two versions instead of one!). If we apply
this technique often enough, we can often replace all applications of a
function by specialized versions, allowing the original function to be
removed (in some cases, this can even give a net reduction of the code
compared to the non-specialized version).
Specialization is useful for our hardware descriptions for functions
that contain arguments that cannot be translated to hardware directly
(polymorphic or higher order arguments, for example). If we can create
specialized functions that remove the argument, or make it translatable,
we can use specialization to make the original, untranslatable, function
obsolete.
\section{Higher order values}
What holds for partial application, can be easily generalized to any
higher order expression. This includes partial applications, plain
variables (e.g., a binder referring to a top level function), lambda
expressions and more complex expressions with a function type (a case
expression returning lambda's, for example).
Each of these values cannot be directly represented in hardware (just like
partial applications). Also, to make them representable, they need to be
applied: function variables and partial applications will then eventually
become complete applications, applied lambda expressions disappear by
applying β-reduction, etc.
So any higher order value will be \quote{pushed down} towards its
application just like partial applications. Whenever a function boundary
needs to be crossed, the called function can be specialized.
TODO: This is section should be improved
\section{Polymorphism}
In Haskell, values can be polymorphic: They can have multiple types. For
example, the function \hs{fst :: (a, b) -> a} is an example of a
polymorphic function: It works for tuples with any element types. Haskell
typeclasses allow a function to work on a specific set of types, but the
general idea is the same.
% A type class is a collection of types for which some operations are
% defined. It is thus possible for a value to be polymorphic while having
% any number of \emph{class constraints}: The value is not defined for
% every type, but only for types in the type class. An example of this is
% the \hs{even :: (Integral a) => a -> Bool} function, which can map any
% value of a type that is member of the \hs{Integral} type class
When generating hardware, polymorphism can't be easily translated. How
many wire will you lay down for a value that could have any type? When
type classes are involved, what hardware components will you lay down for
a class method (whose behaviour depends on the type of its arguments)?
Fortunately, we can again use the principle of specialization: Since every
function application generates separate pieces of hardware, we can know
the types of all arguments exactly. Provided that we don't use existential
typing, all of the polymorphic types in a function must depend on the
types of the arguments (In other words, the only way to introduce a type
variable is in a lambda abstraction). Our top level function must not have
a polymorphic type (otherwise we wouldn't know the hardware interface to
our top level function).
If a function is monomorphic, all values inside it are monomorphic as
well, so any function that is applied within the function can only be
applied to monomorphic values. The applied functions can then be
specialized to work just for these specific types, removing the
polymorphism from the applied functions as well.
By induction, this means that all functions that are (indirectly) called
by our top level function (meaning all functions that are translated in
the final hardware) become monomorphic.
\section{State}
A very important concept in hardware designs is \emph{state}. In a
stateless (or, \emph{combinatoric}) design, every output is a directly and solely dependent on the
inputs. In a stateful design, the outputs can depend on the history of
inputs, or the \emph{state}. State is usually stored in \emph{registers},
which retain their value during a clockcycle, and are typically updated at
the start of every clockcycle. Since the updating of the state is tightly
coupled (synchronized) to the clock signal, these state updates are often
called \emph{synchronous}.
To make our hardware description language useful to describe more that
simple combinatoric designs, we'll need to be able to describe state in
some way.
\subsection{Approaches to state}
In Haskell, functions are always pure (except when using unsafe
functions like \hs{unsafePerformIO}, which should be prevented whenever
possible). This means that the output of a function solely depends on
its inputs. If you evaluate a given function with given inputs, it will
always provide the same output.
TODO: Define pure
This is a perfect match for a combinatoric circuit, where the output
also soley depend on the inputs. However, when state is involved, this
no longer holds. Since we're in charge of our own language, we could
remove this purity constraint and allow a function to return different
values depending on the cycle in which it is evaluated (or rather, the
current state). However, this means that all kinds of interesting
properties of our functional language get lost, and all kinds of
transformations and optimizations might no longer be meaning preserving.
Provided that we want to keep the function pure, the current state has
to be present in the function's arguments in some way. There seem to be
two obvious ways to do this: Adding the current state as an argument, or
including the full history of each argument.
\subsubsection{Stream arguments and results}
Including the entire history of each input (\eg, the value of that
input for each previous clockcycle) is an obvious way to make outputs
depend on all previous input. This is easily done by making every
input a list instead of a single value, containing all previous values
as well as the current value.
An obvious downside of this solution is that on each cycle, all the
previous cycles must be resimulated to obtain the current state. To do
this, it might be needed to have a recursive helper function as well,
wich might be hard to properly analyze by the compiler.
A slight variation on this approach is one taken by some of the other
functional \small{HDL}s in the field (TODO: References to Lava,
ForSyDe, ...): Make functions operate on complete streams. This means
that a function is no longer called on every cycle, but just once. It
takes stream as inputs instead of values, where each stream contains
all the values for every clockcycle since system start. This is easily
modeled using an (infinite) list, with one element for each clock
cycle. Since the funciton is only evaluated once, its output is also a
stream. Note that, since we are working with infinite lists and still
want to be able to simulate the system cycle-by-cycle, this relies
heavily on the lazy semantics of Haskell.
Since our inputs and outputs are streams, all other (intermediate)
values must be streams. All of our primitive operators (\eg, addition,
substraction, bitwise operations, etc.) must operate on streams as
well (note that changing a single-element operation to a stream
operation can done with \hs{map}, \hs{zipwith}, etc.).
Note that the concept of \emph{state} is no more than having some way
to communicate a value from one cycle to the next. By introducing a
\hs{delay} function, we can do exactly that: Delay (each value in) a
stream so that we can "look into" the past. This \hs{delay} function
simply outputs a stream where each value is the same as the input
value, but shifted one cycle. This causes a \quote{gap} at the
beginning of the stream: What is the value of the delay output in the
first cycle? For this, the \hs{delay} function has a second input
(which is a value, not a stream!).
\in{Example}[ex:DelayAcc] shows a simple accumulator expressed in this
style.
\startbuffer[DelayAcc]
acc :: Stream Word -> Stream Word
acc in = out
where
out = (delay out 0) + in
\stopbuffer
\startuseMPgraphic{DelayAcc}
save in, out, add, reg;
% I/O ports
newCircle.in(btex $in$ etex) "framed(false)";
newCircle.out(btex $out$ etex) "framed(false)";
% Components
newReg.reg("") "dx(4mm)", "dy(6mm)", "reflect(true)";
newCircle.add(btex + etex);
in.c = origin;
add.c = in.c + (2cm, 0cm);
out.c = add.c + (2cm, 0cm);
reg.c = add.c + (0cm, 2cm);
% Draw objects and lines
drawObj(in, out, add, reg);
nccurve(add)(reg) "angleA(0)", "angleB(180)", "posB(d)";
nccurve(reg)(add) "angleA(180)", "angleB(-45)", "posA(out)";
ncline(in)(add);
ncline(add)(out);
\stopuseMPgraphic
\placeexample[here][ex:DelayAcc]{Simple accumulator architecture.}
\startcombination[2*1]
{\typebufferhs{DelayAcc}}{Haskell description using streams.}
{\boxedgraphic{DelayAcc}}{The architecture described by the Haskell description.}
\stopcombination
This notation can be confusing (especially due to the loop in the
definition of out), but is essentially easy to interpret. There is a
single call to delay, resulting in a circuit with a single register,
whose input is connected to \hs{outl (which is the output of the
adder)}, and it's output is the \hs{delay out 0} (which is connected
to one of the adder inputs).
This notation has a number of downsides, amongst which are limited
readability and ambiguity in the interpretation. TODO: Reference
Christiaan.
\subsubsection{Explicit state arguments and results}
A more explicit way to model state, is to simply add an extra argument
containing the current state value. This allows an output to depend on
both the inputs as well as the current state while keeping the
function pure (letting the result depend only on the arguments), since
the current state is now an argument.
In Haskell, this would look like \in{example}[ex:ExplicitAcc].
\startbuffer[ExplicitAcc]
-- input -> current state -> (new state, output)
acc :: Word -> Word -> (Word, Word)
acc in (State s) = (State s', out)
where
out = s + in
s' = out
\stopbuffer
\placeexample[here][ex:ExplicitAcc]{Simple accumulator architecture.}
\startcombination[2*1]
{\typebufferhs{ExplicitAcc}}{Haskell description using explicit state arguments.}
% Picture is identical to the one we had just now.
{\boxedgraphic{DelayAcc}}{The architecture described by the Haskell description.}
\stopcombination
This approach makes a function's state very explicit, which state
variables are used by a function can be completely determined from its
type signature (as opposed to the stream approach, where a function
looks the same from the outside, regardless of what state variables it
uses (or wether it's stateful at all).
This approach is the one chosen for Cλash and will be examined more
closely below.
\subsection{Explicit state specification}
We've seen the concept of explicit state in a simple example below, but
what are the implications of this approach?
\subsubsection{Substates}
Since a function's state is reflected directly in its type signature,
if a function calls other stateful functions (\eg, has subcircuits) it
has to somehow know the current state for these called functions. The
only way to do this, is to put these \emph{substates} inside the
caller's state. This means that a function's state is the sum of the
states of all functions it calls, and its own state.
This also means that the type of a function (at least the "state"
part) is dependent on its implementation and the functions it calls.
This is the major downside of this approach: The separation between
interface and implementation is limited. However, since Cλash is not
very suitable for separate compilation (see
\in{section}[sec:prototype:separate]) this is not a big problem in
practice. Additionally, when using a type synonym for the state type
of each function, we can still provide explicit type signatures
while keeping the state specification for a function near its
definition only.
\subsubsection{...}
We need some way to know which arguments should become input ports and
which argument(s?) should become the current state (\eg, be bound to
the register outputs). This does not hold holds not just for the top
level function, but also for any subfunctions. Or could we perhaps
deduce the statefulness of subfunctions by analyzing the flow of data
in the calling functions?
To explore this matter, we make an interesting observation: We get
completely correct behaviour when we put all state registers in the
top level entity (or even outside of it). All of the state arguments
and results on subfunctions are treated as normal input and output
ports. Effectively, a stateful function results in a stateless
hardware component that has one of its input ports connected to the
output of a register and one of its output ports connected to the
input of the same register.
TODO: Example?
Of course, even though the hardware described like this has the
correct behaviour, unless the layout tool does smart optimizations,
there will be a lot of extra wire in the design (since registers will
not be close to the component that uses them). Also, when working with
the generated \small{VHDL} code, there will be a lot of extra ports
just to pass one state values, which can get quite confusing.
To fix this, we can simply \quote{push} the registers down into the
subcircuits. When we see a register that is connected directly to a
subcircuit, we remove the corresponding input and output port and put
the register inside the subcircuit instead. This is slightly less
trivial when looking at the Haskell code instead of the resulting
circuit, but the idea is still the same.
TODO: Example?
However, when applying this technique, we might push registers down
too far. When you intend to store a result of a stateless subfunction
in the caller's state and pass the current value of that state
variable to that same function, the register might get pushed down too
far. It is impossible to distinguish this case from similar code where
the called function is in fact stateful. From this we can conclude
that we have to either:
\startitemize
\item accept that the generated hardware might not be exactly what we
intended, in some specific cases. In most cases, the hardware will be
what we intended.
\item explicitely annotate state arguments and results in the input
description.
\stopitemize
The first option causes (non-obvious) exceptions in the language
intepretation. Also, automatically determining where registers should
end up is easier to implement correctly with explicit annotations, so
for these reasons we will look at how this annotations could work.
TODO: Note about conditions on state variables and checking them.
\subsection{Explicit state annotation}
To make our stateful descriptions unambigious and easier to translate,
we need some way for the developer to describe which arguments and
results are intended to become stateful.
Roughly, we have two ways to achieve this:
\startitemize[KR]
\item Use some kind of annotation method or syntactic construction in
the language to indicate exactly which argument and (part of the)
result is stateful. This means that the annotation lives
\quote{outside} of the function, it is completely invisible when
looking at the function body.
\item Use some kind of annotation on the type level, \eg give stateful
arguments and (part of) results a different type. This has the
potential to make this annotation visible inside the function as well,
such that when looking at a value inside the function body you can
tell if it's stateful by looking at its type. This could possibly make
the translation process a lot easier, since less analysis of the
program flow might be required.
\stopitemize
From these approaches, the type level \quote{annotations} have been
implemented in Cλash. \in{Section}[sec:prototype:statetype] expands on
the possible ways this could have been implemented.
TODO: Say something about dependent types and fixed size vectors
\section[sec:recursion]{Recursion}
An import concept in functional languages is recursion. In it's most basic
form, recursion is a function that is defined in terms of itself. This
usually requires multiple evaluations of this function, with changing
arguments, until eventually an evaluation of the function no longer requires
itself.
Recursion in a hardware description is a bit of a funny thing. Usually,
recursion is associated with a lot of nondeterminism, stack overflows, but
also flexibility and expressive power.
Given the notion that each function application will translate to a
component instantiation, we are presented with a problem. A recursive
function would translate to a component that contains itself. Or, more
precisely, that contains an instance of itself. This instance would again
contain an instance of itself, and again, into infinity. This is obviously a
problem for generating hardware.
This is expected for functions that describe infinite recursion. In that
case, we can't generate hardware that shows correct behaviour in a single
cycle (at best, we could generate hardware that needs an infinite number of
cycles to complete).
However, most recursive hardware descriptions will describe finite
recursion. This is because the recursive call is done conditionally. There
is usually a case statement where at least one alternative does not contain
the recursive call, which we call the "base case". If, for each call to the
recursive function, we would be able to detect which alternative applies,
we would be able to remove the case expression and leave only the base case
when it applies. This will ensure that expanding the recursive functions
will terminate after a bounded number of expansions.
This does imply the extra requirement that the base case is detectable at
compile time. In particular, this means that the decision between the base
case and the recursive case must not depend on runtime data.
\subsection{List recursion}
The most common deciding factor in recursion is the length of a list that is
passed in as an argument. Since we represent lists as vectors that encode
the length in the vector type, it seems easy to determine the base case. We
can simply look at the argument type for this. However, it turns out that
this is rather non-trivial to write down in Haskell in the first place. As
an example, we would like to write down something like this:
\starthaskell
sum :: Vector n Word -> Word
sum xs = case null xs of
True -> 0
False -> head xs + sum (tail xs)
\stophaskell
However, the typechecker will now use the following reasoning (element type
of the vector is left out):
\startitemize
\item tail has the type \hs{(n > 0) => Vector n -> Vector (n - 1)}
\item This means that xs must have the type \hs{(n > 0) => Vector n}
\item This means that sum must have the type \hs{(n > 0) => Vector n -> a}
\item sum is called with the result of tail as an argument, which has the
type \hs{Vector n} (since \hs{(n > 0) => n - 1 == m}).
\item This means that sum must have the type \hs{Vector n -> a}
\item This is a contradiction between the type deduced from the body of sum
(the input vector must be non-empty) and the use of sum (the input vector
could have any length).
\stopitemize
As you can see, using a simple case at value level causes the type checker
to always typecheck both alternatives, which can't be done! This is a
fundamental problem, that would seem perfectly suited for a type class.
Considering that we need to switch between to implementations of the sum
function, based on the type of the argument, this sounds like the perfect
problem to solve with a type class. However, this approach has its own
problems (not the least of them that you need to define a new typeclass for
every recursive function you want to define).
Another approach tried involved using GADTs to be able to do pattern
matching on empty / non empty lists. While this worked partially, it also
created problems with more complex expressions.
TODO: How much detail should there be here? I can probably refer to
Christiaan instead.
Evaluating all possible (and non-possible) ways to add recursion to our
descriptions, it seems better to leave out list recursion alltogether. This
allows us to focus on other interesting areas instead. By including
(builtin) support for a number of higher order functions like map and fold,
we can still express most of the things we would use list recursion for.
\subsection{General recursion}
Of course there are other forms of recursion, that do not depend on the
length (and thus type) of a list. For example, simple recursion using a
counter could be expressed, but only translated to hardware for a fixed
number of iterations. Also, this would require extensive support for compile
time simplification (constant propagation) and compile time evaluation
(evaluation constant comparisons), to ensure non-termination. Even then, it
is hard to really guarantee termination, since the user (or GHC desugarer)
might use some obscure notation that results in a corner case of the
simplifier that is not caught and thus non-termination.
Due to these complications, we leave other forms of recursion as
future work as well.