expression, that adds one to every element of a vector:
\begin{code}
- map ((+) 1) xs
+ map (+ 1) xs
\end{code}
- Here, the expression \hs{(+) 1} is the partial application of the
+ Here, the expression \hs{(+ 1)} is the partial application of the
plus operator to the value \hs{1}, which is again a function that
- adds one to its argument. A lambda expression allows one to introduce an
+ adds one to its (next) argument. A lambda expression allows one to introduce an
anonymous function in any expression. Consider the following expression,
which again adds one to every element of a vector:
\label{img:compilerpipeline}
\end{figure}
-The prototype heavily uses \GHC, the Glasgow Haskell Compiler.
-\Cref{img:compilerpipeline} shows the \CLaSH\ compiler pipeline. As you can
-see, the front-end is completely reused from \GHC, which allows the \CLaSH\
-prototype to support most of the Haskell Language. The \GHC\ front-end
-produces the program in the \emph{Core} format, which is a very small,
-typed, functional language which is relatively easy to process.
-
-The second step in the compilation process is \emph{normalization}. This
-step runs a number of \emph{meaning preserving} transformations on the
-Core program, to bring it into a \emph{normal form}. This normal form
-has a number of restrictions that make the program similar to hardware.
-In particular, a program in normal form no longer has any polymorphism
-or higher order functions.
-
-The final step is a simple translation to \VHDL.
+The output of the \GHC\ front-end is the original Haskell description
- translated to \emph{Core}~\cite{Sulzmann2007}, which is smaller, functional,
- typed language that is relatively easier to process than the larger Haskell
++translated to \emph{Core}~\cite{Sulzmann2007}, which is smaller, typed,
++functional language that is relatively easier to process than the larger Haskell
+language. A description in \emph{Core} can still contain properties which have
+no direct translation to hardware, such as polymorphic types and
+function-valued arguments. Such a description needs to be transformed to a
+\emph{normal form}, which only contains properties that have a direct
+translation. The second stage of the compiler, the \emph{normalization} phase
+exhaustively applies a set of \emph{meaning-preserving} transformations on the
+\emph{Core} description until this description is in a \emph{normal form}.
+This set of transformations includes transformations typically found in
+reduction systems for lambda calculus~\cite{lambdacalculus}, such a
+$\beta$-reduction and $\eta$-expansion, but also includes self-defined
+transformations that are responsible for the reduction of higher-order
+functions to `regular' first-order functions.
+
+The final step in the compiler pipeline is the translation to a \VHDL\
+\emph{netlist}, which is a straightforward process due to resemblance of a
+normalized description and a set of concurrent signal assignments. We call the
+end-product of the \CLaSH\ compiler a \VHDL\ \emph{netlist} as the resulting
+\VHDL\ resembles an actual netlist description and not idiomatic \VHDL.
\section{Use cases}
+
+\subsection{FIR Filter}
\label{sec:usecases}
As an example of a common hardware design where the use of higher-order
functions leads to a very natural description is a FIR filter, which is
The \hs{zipWith} function is very similar to the \hs{map} function seen
earlier: It takes a function, two vectors, and then applies the function to
each of the elements in the two vectors pairwise (\emph{e.g.}, \hs{zipWith (*)
-[1, 2] [3, 4]} becomes \hs{[1 * 3, 2 * 4]} $\equiv$ \hs{[3,8]}).
+[1, 2] [3, 4]} becomes \hs{[1 * 3, 2 * 4]}).
-The \hs{foldl1} function takes a function, a single vector, and applies
+The \hs{foldl1} function takes a binary function, a single vector, and applies
the function to the first two elements of the vector. It then applies the
-function to the result of the first application and the next element from
-the vector. This continues until the end of the vector is reached. The
-result of the \hs{foldl1} function is the result of the last application.
-As you can see, the \hs{zipWith (*)} function is pairwise
-multiplication and the \hs{foldl1 (+)} function is summation.
-
-Returning to the actual FIR filter, we will slightly change the
-equation belong to it, so as to make the translation to code more obvious.
-What we will do is change the definition of the vector of input samples.
-So, instead of having the input sample received at time
-$t$ stored in $x_t$, $x_0$ now always stores the current sample, and $x_i$
-stores the $ith$ previous sample. This changes the equation to the
-following (Note that this is completely equivalent to the original
-equation, just with a different definition of $x$ that will better suit
-the transformation to code):
+function to the result of the first application and the next element in the
+vector. This continues until the end of the vector is reached. The result of
+the \hs{foldl1} function is the result of the last application. It is obvious
- that the \hs{zipWith (*)} function is basically pairwise multiplication and
- that the \hs{foldl1 (+)} function is just summation.
++that the \hs{zipWith (*)} function is pairwise multiplication and that the
++\hs{foldl1 (+)} function is summation.
+
+Returning to the actual FIR filter, we will slightly change the equation
+describing it, so as to make the translation to code more obvious and concise.
+What we do is change the definition of the vector of input samples and delay
+the computation by one sample. Instead of having the input sample received at
+time $t$ stored in $x_t$, $x_0$ now always stores the newest sample, and $x_i$
+stores the $ith$ previous sample. This changes the equation to the following
+(note that this is completely equivalent to the original equation, just with a
+different definition of $x$ that will better suit the transformation to code):
\begin{equation}
y_t = \sum\nolimits_{i = 0}^{n - 1} {x_i \cdot h_i }
\end{equation}
-Consider that the vector \hs{hs} contains the FIR coefficients and the
-vector \hs{xs} contains the current input sample in front and older
-samples behind. The function that shifts the input samples is shown below:
+The complete definition of the FIR filter in code then becomes:
\begin{code}
-x >> xs = x +> tail xs
+fir (State (xs,hs)) x = (State (x >> xs,hs), xs *+* hs)
\end{code}
-Where the \hs{tail} function returns all but the first element of a
-vector, and the concatenate operator ($\succ$) adds a new element to the
-front of a vector. The complete definition of the FIR filter then becomes:
+Where the vector \hs{hs} contains the FIR coefficients and the vector \hs{xs}
+contains the latest input sample in front and older samples behind. The code
+for the shift (\hs{>>}) operator that adds the new input sample (\hs{x}) to
+the list of previous input samples (\hs{xs}) and removes the oldest sample is
+shown below:
\begin{code}
-fir (State (xs,hs)) x = (State (x >> xs,hs), xs *+* hs)
+x >> xs = x +> init xs
\end{code}
-The resulting netlist of a 4-taps FIR filter based on the above definition
-is depicted in \Cref{img:4tapfir}.
+The \hs{init} function returns all but the last element of a vector, and the
- concatenate operator ($\succ$) adds a new element to the left of a vector. The
++concatenate operator ($\succ$) adds a new element to the front of a vector. The
+resulting netlist of a 4-taps FIR filter, created by specializing the vectors of the above definition to a length of 4, is depicted in \Cref{img:4tapfir}.
\begin{figure}
\centerline{\includegraphics{4tapfir.svg}}