Lava is an experimental system design to aid the digital design of circuits by providing a powerful library for

**composing**structural circuit descriptions. Lava is designed to help specify the**layout**of circuits which can improve performance and reduce area utilization. An implementation of Lava for Xilinx's FPGAs can be found at Hackage and installed with "cabal install xilinx-lava".
The following pages give a flavor of the Lava HDL and how it can be used to describe circuits for implementation of Xilinx's Virtex family of FPGAs. These pages assume a good understanding of Xilinx's Virtex FPGA architecture and of the Haskell lazy functional programming language. The number of people that know about both can easily fit inside a medium sized elevator. So it may be useful to read A Gentle Introduction to Haskell if you are unfamiliar with functional programming languages. Information about Xilinx's FPGA can be found at http://www.xilinx.com. A detailed set of Lava tutorials are also available.

## Describing Netlists in Lava

Structural netlists are described in Lava by composing functions that represent basic library elements or composite circuits. This is very similar to the way the structural netlists are described in VHDL or in Verilog which rely on netlist naming to compose circuit elements. Later we present combinators which provide a much more powerful way to combine circuits.

Most hardware description languages provide a library of basic gates that correspond to the Xilinx Unified Library components. Designers can build up netlists by creating instances of these abstract gates and writing them together. The implementation software maps these gates to specific resources on the FPGA. Most logic functionality is mapped into the slices of CLBs. The top half of a Virtex-II slice is shown below:

Lava provides a more direct way to specify the mapping into CLB, slices and LUTs. Just like the Xilinx Unified Library the Xilinx Lava library contains specific functions (circuits) that correspond to resources in a Virtex slice e.g. muxcy and xorcy. However, instead of trying to enumerate every possible one, two, three and four input function that can be realized in a LUT Lava instead provides a higher order function for creating a LUT configuration from a higher level specification of the requited function.

As an example consider the task of configuring a LUT to be a two input AND gate. This can be performed by using the lut2 combinator with an argument that is the Haskell function that performs logical conjunction written as &.

```
and2 :: (Bit, Bit) -> Bit
and2 =
```**lut2** (&)

The

**lut2**higher order combinator takes any function of two Boolean parameters and returns a circuit that corresponds to a LUT programmed to perform that function. A circuit in Lava has a type that operates over structures of the Bit type. The and2 circuit is defined to take a pair of bits as input and return a single bit as its output.
Note that

**lut2**is a**higher order**combinator because it takes a function as an argument (the & function) and returns a function as a result (the circuit which ANDs its two inputs). Indeed the**lut2**combinator can take any function that has the type Bool -> Bool -> Bool and it returns a circuit of type (Bit, Bit) -> Bit. The type of the**lut2**function is similar to:lut2 :: (Bool -> Bool -> Bool) -> ((Bit, Bit) -> Bit)

When Lava generates a VHDL or EDIF netlist for this component it will instance a LUT with the appropriate programming information. Here is what the generated VHDL for the and2 gate might look like:

lut2_1 : lut2 generic map (init => "1000") portmap (i0 => lava(5), i1 => lava (6), o => lava(4)) ;

When realized in the top part of a slice as shown in the picture above this gate would be mapped to the upper G function generator which will be configured as a LUT.

As another example consider the definition of a three input AND-gate. First, we define a function in Haskell that performs the AND3 operation:

and3function :: Bool -> Bool -> Bool -> Bool and3function a b c = a & b & c

The first line gives the type of and function as something that takes three boolean values and returns a boolean value. The next line defines the and3function in terms of the Haksell binary & function. Now we are in a position to define a circuit that ANDs three inputs:

`and3 = lut3 and3function`

The general idea is that instead of trying to provide a library that tries to enumerate some of the one, two, three and four input logic functions Lava provides four combinators

**lut1**,**lut2**,**lut3**and**lut4**. These combinators take functions expressed in the embedding language (Haskell) and return LUTs that realize these functions. This makes it convenient to express exactly what gets packed into one LUT. Later we will see that these higher order combinators are actually overloaded which allows LUT contents to be specified in many different kinds of ways (including an integer or a bit-vector).
Lava does provide a module that contains definitions for commonly used gates mapped into LUTs. These include:

inv = lut1 not and2 = lut2 (&) or2 = lut2 (||) xor2 = lut2 exor muxBit sel (d0, d1)= lut3 muxFn (sel, d0, d1)

assuming the following functions are available in addition to the Haskell prelude:

exor :: Bool -> Bool -> Bool exor a b = a /= b muxFn :: Bool -> Bool -> Bool -> Bool muxFn sel d0 d1 = if sel then d1 else d0

To make a netlist that corresponds to a NAND gate one could simply perform direct instantiation and wire up the signals appropriate just like in VHDL or Verilog:

```
nandGate (a, b) = d
where
d = inv c
c = and2 (a, b)
```

Lava generates a netlist containing two LUT instantiations for this circuit: a LUT2 to implement the AND gate and a LUT1 to implement the inverter. How many LUTs are used to realize this circuit on the FPGA? If no information is given about the location of these two circuits then the mapper is free to merge both the LUT2 for the AND gate and the LUT1 for the inverter into one LUT. However, if the inverter has been laid out in a different location from the AND gate then the two components will not be merged. By default each gate is at logical position (0.0). Since the description above does not translate the two sub-circuits to another location they are understood to occupy the same function generator location and will be merged into one LUT.

A partial list of some of the basic Virtex slice resources supported by Lava is shown below:

gnd :: Bit vcc :: Bit fd :: Bit -> Bit -> Bit fde :: Bit -> Bit -> Bit -> Bit muxcy :: (Bit, (Bit, Bit)) -> Bit muxcy_l :: (Bit, (Bit, Bit)) -> Bit xorcy :: (Bit, Bit) -> Bit xorcy_l :: (Bit, Bit) -> Bit muxf5 :: (Bit, (Bit, Bit)) -> Bit muxf6 :: (Bit, (Bit, Bit)) -> Bit muxf7 :: (Bit, (Bit, Bit)) -> Bit muxf8 :: (Bit, (Bit, Bit)) -> Bit

## Layout in Lava

### The Lava Coordinate System

Lava uses a Cartesian coordinate system with the origin at the bottom left hand corner of the screen. A Lava design can have many independent coordinate systems and each of these will be represented by an H_SET in the implementation netlist. The Lava coordinate system is designed to make it easier to produce circuit descriptions that are portable between the Virtex and Virtex-II architectures by abstracting as much as possible the CLB and slice boundaries.

Circuits are conceptually enclosed in rectangular tiles which can be laid out relative to each other. There are two kinds of tiles that Lava supports: (a) two-sided tiles which communicate only along their left and right hand sides and (b) four sided tiles that can communicate along all four sides of the tile.

### Serial Composition

The serial composition combinator in Lava is written as

**>->**and it takes two circuits as its arguments and returns a new composite circuit. This combinator performs two distinct functions:- Lava composes behavior by taking the output of the first circuit and connecting it to the input of the second circuit (i.e. like mathematical functional composition but written the other way around).
- Lava places the second circuit to the right of the first circuit. All circuits start off with their bottom left hand corner at location (0,0). By using layout combinators the relative locations of circuits can be modified.

An example of

**serial composition**is shown below for the circuit and2 >-> inv. The circuit and2 takes a pair of signals as its input and returns a single output bit. This is illustrated in the diagram below:
The and2 circuit is realized in a two-sided Lava tile which means that signals (either inputs or outputs) can occur on the left or right hand sides of the tile but not the top or bottom. The inv circuit is also realized in a two-sided tile. To make an NAND gate form these components all one needs to do is to connect the output of one circuit to the input of the other circuit. This task can be performed by the serial composition combinator without having to name the intermediate signals.

However the serial composition combinator does more than just connect wires together. Serial composition looks at the sizes of the circuits it has to compose and it lays out one circuit directly next to the other circuit with their bottoms aligned. For example consider the tiles containing the inv and and2 gates which which each have size (1,1). The serial composition combinator takes the output of the and2 gate and connects it to the input of the inv gate resulting in a circuit that behaves like a NAND-gate. It also places the inv gate to the right of the and2 gate so it ends up in location (1,0). The composite NAND-gate circuit tile has size (2,1) as shown below.

When Lava generates an EDIF netlist it translates its own coordinate system into an appropriate coordinate system for the chosen target architecture. For Virtex family devices it will generate relative location attributes using CLB and slice coordinates. The and2 gate at Lava location (0,0) will be mapped to R0C0.S1 (the left hand slice of a CLB at relative location (0,0)). The inv gate will be mapped from the Lava location (1,0) to R0C0.S0 (the right hand slice of a CLB at relative location (0,0)). The absolute location of the circuit on the FPGA is left to the Xilinx placer. The corresponding layout on a Virtex XCV300 FPGA is shown below (captured from the Xilinx FPGA Floorplanner):

The and2 gate at Lava location (0,0) has been mapped into the upper LUT of the left hand slice of the CLB at R32C1 i.e. R32C1.S1 (the left hand corner of this Virtex device). The inv component has been mapped into the upper LUT of the right hand slice in the same CLB i.e. R32C1.S0. Note that within the slice the Xilinx tools are free to map to either the upper or lower LUT (F or G). Lava coordinates can specify which logic should be mapped into a slice by requesting either F (y coordinate an even number) or G (y coordinate an odd number) in a

**relative**coordinate system but the Xilinx mapper will decide the actual mapping and absolute location. For the Virtex-II architecture Lava generates slice-based coordinates.**Parallel Composition**

The serial composition combinator is useful when we wish to compose and lay out two circuits that communicate. Sometimes however we wish to lay out circuits relative to each other which do not communication. One of the many combinators that Lava provides for composing circuits in

**parallel**is the**par2**combinator. This combinator takes two two sided tiles and placed the second tile above the first tile aligned along their left hand edges. An example layout for the circuit**par2**inv and2 is shown below:
Any circuit or combinator that takes two inputs can be written as in infix operator in Lava by enclosing the name of the circuit or combinator in back quotes as shown above i.e.

**par2**inv and2 is the same as inv `**par2**` and2. The**par2**combinator produces a composite circuit that takes a two element tuple as its input and returns a two element tuple. The first element of the input tuple is the input destined for the first (lower) circuit and the second element of the input tuple is the input destined for the second (upper) circuit. The first element of the output tuple is the output from the first (lower) circuit and the second element of the output tuple is the output from the second (upper) circuit. The layout of the composite circuit tile of size (1,2) in the Lava coordinate system is shown below.
The

**par2**combinator operates over just two circuits. Three or more circuits can be composed by using**par2**as an infix operator but this results in circuits with awkward tuple types. The**par**combinator takes a list of circuits (all of the same type) and lays them out from the bottom to the top:
Sometimes it is useful to apply the same circuit to every element of an input bus. We can use the Haskell higher order map function and the Lava

**par**combinator to define a circuit combinator which performs exactly this task. The Lava combinator that maps a circuit across a bus is called**maP**(with a capital P to it does not clash with the Haskell function map) and is defined as:**maP** r = **par** (repeat r)

This combinator takes any circuit r (which is considered to be a two sided tile) and replicates it as many times as there are elements in the input bus and then layouts out these replicated circuits vertically. For example to invert every bit on a bus one would write

**maP**invresulting in the layout shown below:
Another use of maP is to vertically place registers across each element of a bus. The basic register in Lava that corresponds to FD takes two inputs: the clock and the D input signal. The circuit vreg registers each element of a bus and is define as:

vreg clk =maP(fd clk)

### Serial Overlay Combinator

A register fits in a tile of size (1,1). To describe an AND gate with a register at the output we could use serial composition:

and2Reg clk = and2 >-> fd clk

This produces the following layout:

Why is the register realized in the adjacent slice and not the slice containing the function generator for the AND2 gate? This is because the and2 circuit occupies a tile of size (1,1) and the register also occupies a tile of size (1,1) and when the serial composition combinator is asked to put one tile next to the other that is exactly what it does. This results in the register being realized in location (1,0). If this is not what is wanted then a variant of the serial composition combinator can be used to compose two circuits without translating the second circuit. This combinator is called the

**serial overlay combinator**and is written as**>|>**. The circuit and2**>|>**fd clk will realize both the and2 circuit and the inv circuit into location (0,0) giving:
Now the and2 gate and the FD register are realized in the top half of one slice. Lava will always map and place circuits exactly in the way specified by the layout combinators.

### Describing Non Left-to-Right Dataflow

The

**>->**serial composition combinator only composes circuits with a left to right data-flow. There are three other serial compositions combinators that describe right to left (<-<), bottom to top (/\) and top to bottom (\/) serial composition. They are illustrated below:## Four Sided Tiles

**Beside and Below**

To describe circuits that communicate on more than two faces of a tile Lava also provides support for four sided tiles. A circuit can be put inside a four sided tile if it has a type which takes a two element tuple as its input and returns a two element tuple as shown below:

A four sided tile is divided into the input signal and output signal portions by the gray diagonal line. The input to a circuit that fits into a four sided tile must be a tuple (a, b) where the a signal is connected to the bottom of the tile and the b signal is connected to the left of the tile. These signals can be single bits or composite signals. The result type of a circuit that fits into a four sided tile must also be a tuple (c, d) where the c signal is connected to the right hand side of the tile and the d signal is connected to the top of the tile.

To put one four sided tile below another four sided tile requires a new combinator because these tiles do not have the appropriate types for the

**par2**combinator. Lava provides the below combinator for this task and the picture below shows the circuit r**`below`**s:
The circuit r has the type (e,b) -> (c, x) and the circuit s has the type (x,e) -> (f, g) and the composite circuit has the type (e, (b, e)) -> ((c, f), g) where x is the type of the intermediate signal that r and s communicate along. There is also a

**beside**combinator that lays out s to the right of r.**Row and Col**

Four sided tiles can be replicated and composed to form columns and rows. The

**col**combinator tiles a circuit r vertically and forms connections between the top and bottom faces of the tile:
The row combinator composes tiles horizontally from left to right and is based on the

**beside**combinator. Rows and column combinators can be used to describe many interesting layout patterns and emulate layout grids which are not four sided. Here is how a hex-connected grid might be described in terms of four sided tiles:## An Adder Example

### A Ripple Carry Adder Example (using the Virtex Carry Chain)

This section describes how to design a ripple-carry adder in Lava that uses the Virtex carry chain. This component is widely used in many other Lava examples. First we present how to describe a one-bit adder cell using the regular netlist style. Then we use layout combinators to create a vertical column of these cells that make up a carry-chain ripple-carry adder.

**A One-Bit Adder**

A one-bit adder can be accommodated in one half of a slice using just one function generator as a LUT, one MUXCY and one XORCY. For this reason we will describe the one-bit adder in netlist style. A schematic for a one-bit carry chain adder is shown below.

A column layout combinator will be used later to tile several one-bit adders vertically to make an

*n*-bit adder. Before we tile the one-bit adder vertically we have to carefully consider which sides of the tile the signals are connected to. We shall assume a left to right data flow with the two bits to be added connected to the left hand side of the tile and the sum bit appearing from the right hand side of the tile. Since the carry must flow upwards in the Virtex architecture the cin signal must come into the bottom of the tile and the cout signal must emerge from the top of the tile. This gives the following four-sided tile orientation:
This four-sided tile represents a circuit with input ((a,b), cin) because (a,b) is the left input and cin in the bottom input. This tile also has the output (sum, cout) because sum is the right output and cout is the top output. This fixes the type for this one-bit adder adder tile which can now be described as a Lava netlist as follows:

```
oneBitAdder :: (Bit, (Bit, Bit)) -> (Bit, Bit)
oneBitAdder (cin, (a,b))
= (sum, cout)
where
part_sum = xor2 (a, b)
sum = xorcy (part_sum, cin)
cout = muxcy (part_sum, (a, cin))
```

This is very similar to what one would write in VHDL or Verilog except that the instance names are anonymous. Furthermore, the Lava description includes information in the shape of the types that indicates which faces of the tile signals occur and in what order. This is essential in order to compose tiles using combinators which automatically glue together circuit interfaces and layout out circuits in specific patterns.

**A 4-bit Carry Chain Adder**

To make a 4-bit adder all we need to do is to tile vertically four one-bit adder tiles:

In a conventional HDL this can be described either by copying and pasting the instantiating code for each one-bit adder and then naming intermediate nets which cause components to be connected together to form the composite four-bit adder netlist. Constructs like

**for**...**generate**in VHDL help to capture such repetition more concisely but internal nets still have to be named and there is no information about layout. The Lava description of the circuit show above is simply:**col** 4 oneBitAdder

This is much more concise than equivalent conventional HDL descriptions. It also includes information about layout: the first one-bit adder tile is at the bottom and then the other are stacked on top of it in sequence (RLOCs are generated to enforce this layout). Internal nets are anonymous and circuits are composed automatically by taking the input from below and attaching it to the carry input of the tile.

An example Virtex layout of a 8-bit adder generated by the expression

**col**8 oneBitAdder is show below:
In Lava a parameterized n-bit adder can be defined as:

```
adder :: Int -> (Bit, ([Bit], [Bit])) -> ([Bit], Bit)]
adder n (cin, (a, b))
= col n oneBitAdder (cin, zip a b)
```

This parameterized adder function uses an integer parameter n to determine the size of the adder. This controls how many one-bit adders are placed in a column. Each one-bit adder is expecting a par of bits on its left edge. To form the correct input for a column of one-bit adders we have to pair-wise associate the two inputs vectors a and b. This is accomplished by using the list processing function zip which has the Haskell type [a] -> [b] -> [(a, b)]. For example:

zip [1, 4, 9] [2, 13, 5] = [(1,2), (4,13), (9,5)]

The use of zip in adder ensures that the nth bit of a is added to the

*n*th bit of b.
A specialized version of the adder which has no carry in or carry out can defined as:

```
adderNoCarry :: Int -> ([Bit], [Bit])) -> [Bit]
adderNoCarry n (a,b)
= sum
where
(sum, carryOut) = adder n (gnd, (a,b))
```

This definition uses the gnd component to provide the carry in signal and just discards the carry out.

A registered 4-bit adder can be made by composing the adderNoCarry component and the vreg component using >|> (the serial overlay operator) to make sure the adder and register share the same location:

A parameterized

*n*-bit registered adder can be defined as:registeredAdder :: Int -> Bit -> ([Bit], [Bit]) -> [Bit] registeredAdder n clk = adderNoCarry n >|> vreg clk

The Virtex layout produced for registeredAdder 4 is shown below.

##
**An Adder Tree in Lava**

### A Carry Chain Adder Example

This section describes how to build an adder tree using the carry chain adders presented in the previous section.

### A Recursive Layout

Consider the task of adding eight numbers n1, n2, .. n8 use a binary tree of adders as shown below.

In a language like VHDL one can write a recursive function that takes an unconstrained array of numbers as input and returns their sum computed with an adder tree. However, there is no standard way in VHDL of specifying how the adder tree should be laid out. Adder trees are typically pipelined and it takes just one adder to be badly placed to degrade the performance of the whole adder tree. Since Lava can specify how the adders are to be laid out the design engineer can make decisions and calculations about intermediate wire delays. One example layout scheme for a pipelined adder tree is shown below (the clock wires are now shown):

This layout scheme places the adders in a row with the final addition occurring in the middle. Both the left and right hand sides of the final adder also contain adder trees in which the final sum occurs in the middle. This recursive structure continues until the leaf additions are encountered (e.g. the adder for n1 and n2).

### The middle Combinator

To help describe the idea of three components laid out horizontally with the middle component receiving its inputs from its neighbors we introduce the middle combinator middle r s t which gives the following layout:

Note that this combinator produces a two-sided tile which has two inputs on the left and one output on the right.

### Tree Circuits in Lava

Using the middle combinator we can define a tree circuit combinator for any kind of two input and one output circuit:

tree circuit [a] = a tree circuit [a, b] = circuit (a, b) tree circuit xs = middle (tree circuit) circuit (tree circuit)(halve xs)

- The first line of this definition is used when a singleton list is given to the tree circuit. In this case the result is just the sole element of the list.
- The second line is used when a two element list is used with a tree circuit. In this case the result is obtain by processing the two inputs with the circuit that is the first parameter of the tree combinator.
- In the third line case the middle circuit will be the parameterized two input and one output circuit and the left and right circuits will be recursively defined trees of this circuit. The input is halved with the first half going to the left sub-tree and the second half going to the right sub-tree.

A combinational adder tree with bit-growth can be made with the flexibleAdder circuit. This circuit has the type (Bit, Bit) -> (Bit) so it is suitable for use as an element of an adder tree. To define a combinational adder tree one simply writes:

adderTree = tree flexibleAdder

It is possible to rewrite a call to the tree combinator to see how it works for a particular circuit and input list:

treeflexibleAdder [i1, i2, i3, i4] = (middle(treeflexibleAdder) circuit (treeflexibleAdder)) (halveList [i1, i2,i3, i4]) = (middle(treeflexibleAdder) circuit (treeflexibleAdder)) [[i1, i2], [i3, i4]] = flexibleAdder (treeflexibleAdder [i1, i1],treeflexibleAdder [i2, i3]) = flexibleAdder (i1+i2, i3+i4) = i1+i2+i3+i4

We can use the flexibleAdderFD circuit to make a pipelined adder tree:

adderTreeFD clk =

**tree**(flexibleAdderFD clk)
The adders in this tree will use FD register components. An example layout of four 96-input pipelined adder trees stacked upon each other is shown below on a XCV300 part. The bit-growth in the adder tree helps to identify the recursive nature of the layout.

The adder trees shown above sum 96 9-bit numbers each and with this layout the circuit operates at 173MHz on a XCV300 at speed grade 6. If the layout information is removed and the same netlist is placed and routed again the maximum speed is 129MHz i.e. 34% slower. The graph below shows that specifying layout had significant performance advantages. The blue bars represent the speed of circuits with layout information and the red bars represented the speed of the same netlist with the layout information (RLOCs) removed. Each pair of bars corresponds to a target timing specification. The leftmost pair of bars is for the case where the place and route tools were asked to meet a timing specification of 180MHz but delivered 173MHz for the RLOC-ed circuit and 129MHz without the RLOCs.

Providing layout information also speeds out the place and route times as shown in the graph below. For the 180MHz timespec case the netlist with layout took less then 400 seconds to route (nothing needs to be placed) but took more than 1800 seconds to place and route without layout information.

The FPGA Editor view of the four placed adder trees is shown below.

Without the RLOCs the following implementation is produced:

The adder tree layout was easy to express in Lava and has produced significant performance advantages. Many other tree layout schemes can be easily explored in Lava. It is also very easy to make trees of other kinds of circuits like multiplier etc. All one has to do is to supply such circuits to the first argument of the

**tree**combinator.#
**A Constant Coefficient Multiplier (KCM) Core in Lava**

### A Constant Coefficient Multiplier Example in Lava

This section describes how to build a constant coefficient multiplier (KCM) for the Virtex family of FPGAs using a combination of table lookups and additions. There are many other kinds of KCM: this one has been chosen to help illustrate how to design with distributed memory components in Lava.

The architecture of this multiplier is based on performing many 4-bit (or smaller) multiplications and then combining the results of these partial products with a weighted adder tree. Optionally the adder tree can be pipelined. The reason for performing 4-bit multiplications is that these can be performed quickly with lookup tables (since each LUT has four inputs). A specific multiplier generated by the Lava KCM core is shown below.

This KCM multiplies a 11-bit signed dynamic input A = a10 ..0 by a 11-bit constant coefficient K. The multiplication AK is performed by composing the results of several four-bit multiplications by exploiting the identity:

AK = -28 a10 ..8 K + 24 a7 ..4 K + a3 ..0 K

The rightmost table multiplies the lower four bits of the inputs by the constant coefficient i.e. X = a3 ..0 K and the middle table computes Y = a7 ..4 K where both a3 ..0 and a7 ..4 are treated as unsigned numbers. The leftmost table calculates Z = (a10, a10, a9, a8) Kwhere (a10, a10, a9, a8) is treated as a sign-extended bit-vector. To calculate the final results the partial products Y and Z have to be appropriately weighted before addition A weighted adder tree is used to compose the partial products and the size of the intermediate adders is reduce by reading off directly the bottom four bits of the partial product X and the result of adding the remaining bits of X to Y. For the pipelined version of this multiplier Lava places registers on the intermediate wires making sure to balance the delays.

The four-bit multiplications are performed by table lookups using distributed memory (i.e. LUTs). In this case an 11 bit constant must be multiplied by a four bit dynamic input which requires a table with a 15-bit output. This can be easily made by using 15 four-input LUTs all with the same input. To avoid unwanted mapper optimizations we use the ROM16X1 component to create a LUT with specific lookup table contents.

Lava provides a family of functions that make it easy to create tables made out of ROM16X1 components. One of the functions available from the Lava LUTROM module is rom16x which will create a four-input n-output table from data specified as a list of numbers.

lut16x :: [Int] -> (Bit, Bit, Bit, Bit) -> Bit

The code for a four-bit unsigned KCM in Lava is shown below.

```
unsignedFourBitKCM :: Int -> [Bit] -> [Bit]
unsignedFourBitKCM coeif addr
= rom16x maxwidth multiplication_results padded_addr
where
padded_addr = padAddress addr
nr_addrs = length addr
multiplication_results
= pad_width 0 16 [coeif * i | i <- [0..2^nr_addrs-1]]
maxwidth
= maximum
(map unsignedBitsNeeded multiplication_results)
```

The first parameter to this function is coeif which is the constant coefficient. The second argument is a bit-vector which is represented as a list in Lava (with the LSB as the first element). To allow multiplications by bit-vectors less then four bits in width this function pads out the table with zero values for smaller multiplications. This is done by first forming a table of the multiplication results i.e.

[coeif * i | i <- [0..2^nr_addrs-1]]

which for an address less than four bits will have less than 16 elements. Then the pad_width function is used to add zero elements to end of the list to bring the length of the list up to 16. The result is bound to the variable name multiplication_results which is examined to find the largest product which in turn is used to determine how many output bits are required (maxwidth). Now the product can be computed by by forming a 16 element table with maxwidth output bits using the rom16x function. A singed four-bit KCM can be built in a similar manner (this time using sign extension on the input bits).

To make an unsigned KCM the following steps are taken:

- The input bus is chopped into groups of 4-bits. This is achieved using the chop list function.
- Each of these 4-bits is multiplied by the constant coefficient using the unsigedFourBitKCM circuit. This can be easily done by the expression
**maP**(unsignedFourBitKCM coeif) which places the 4-bit KCMs on top of each other. - The results of each of these partial products has to be suitably weighted to allow them to be combined. This is done by zipping the partial products bus with the list [20, 24, 28...]
- A weighted adder tree is used to sum the partial products. Combinational and pipelined KCMs will have different adder trees but we would like to abstract over this difference so that one generic unsigned KCM function needs to be written. This can be done by passing the required adder tree circuit as a higher order circuit.
- The result of the weighted adder tree circuit is a pair which has 20 as the first element and the final product as the second element. Since we only want the product we just use the
**snd**function to project it out.

Each of the stages described above is composed in series as shown below for the case of a 12-bit dynamic input:

Since each unsignedFourBitKCM is long and thin it is better to lay them out horizontally by using the horizontal version of the

**maP**combinator called**hmaP**:
This picture leads to the Lava definition of unsignedKCM shown below:

```
unsignedKCM fourBitKCM adderTree coeif
= chop 4 >->
hmaP (fourBitKCM coeif) >->
insertWeights >->
adderTree >->
snd
```

Note how the specific four-bit KCM circuit to be used is passed as a higher order parameter. The unsignedKCM higher order circuit can be instantiated with a specific four-bit KCM to make either combinational or sequential KCMs. Similarly the adderTree circuit is passed as a higher order circuit. To make a specific unsigned KCM we also need to specify exactly which adder tree circuit to use. If we used a combinational weighted tree the resulting KCM would be a combinational circuit. To help design a weighted adder we define a type that describes the "virtual" signals that flow along the wires of a weighted adder. The type WtNr is defined to be a tuple which contains a weight represented as an integer and a bit-vector. These signals are called "virtual" because the weight component is an integer which is only used during elaboration and does not correspond to a wire in the circuit.

type WtNr = (Int, [Bit])

The unsignedWeightedAdder circuit takes a pair of such weighted numbers and returns a weighted number. This is expressed by the following type signature:

unsignedWeightedAdder :: (WtNr, WtNr) -> WtNr

This circuit assumes the first element of the input pair has a larger weight than the second element. If this is not the case this circuit calls itself again with the elements of the pairs swapped. This is achieved by using a pattern match guard to ensure that the first input has a higher weight than the second input:

unsignedWeightedAdder ((weight1 ,a1), (weight2,a2)) | weight1 < weight2 = unsignedWeightedAdder ((weight2, a2), (weight1,a1))

Otherwise unsignedWeightedAdder computes the difference between the two weights to compute how many of the lower bits of the second input to pass through. The remaining bits are then added to the first input to compute the upper bits of the final sum. The weight of the result is the higher input weight and the resulting addition is formed by composing the bit-vector from the addition with the lower bits.

```
unsignedWeightedAdder ((weight1,a1), (weight2,a2))
= (weight2, lower_bits ++ part_sum)
where
weight_difference = weight1 - weight2
lower_bits = take weight_difference a2
a2Upper = drop weight_difference a2
part_sum = flexibleUnsignedAdderNoGrowth (a1, a2Upper)
```

Putting these lines of code together produces the complete code for the unsignedWeightedAdder circuit:

```
unsignedWeightedAdder :: (WtNr, WtNr) -> WtNr
unsignedWeightedAdder
((weight1 ,a1), (weight2,a2)) | weight1 < weight2
= unsignedWeightedAdder ((weight2, a2), (weight1,a1))
unsignedWeightedAdder ((weight1,a1), (weight2,a2))
= (weight2, lower_bits ++ part_sum)
where
weight_difference = weight1 - weight2
lower_bits = take weight_difference a2
a2Upper = drop weight_difference a2
part_sum = flexibleUnsignedAdderNoGrowth (a1, a2Upper)
```

To make a specific unsigned combinational KCM we just need to give a tree of unsigned weighted adders to the unsignedKCM circuit:

unsignedCombinationalKCM = unsignedKCM unsignedFourBitKCM (tree unsignedWeightedAdder)

To make a pipelined KCM we need to provide the unsignedKCM circuit with a pipelined version of the four-bit KCM circuit and a pipelined version of the weighted adder tree circuit. A pipelined version of the four-bit KCM just places a register at the output of the lookup tables with a C clock input and a CE clock enable input. This circuit is called unsignedFourBitKCMCE and is easily defined as:

unsignedFourBitKCMCE clk ce coeif = unsignedFourBitKCM coeif >|> vregE clk ce

i.e. a combinational four-bit unsigned KCM serially composed (by overlaying) with a register-bus with clock and clock enable inputs. An unsigned weighted adder tree can be made by making a tree of unsigned weighted adders that have their outputs registered in a similar manner. This allows a registered unsigned KCM to be defined as:

unsignedRegisteredKCM clk ce = unsignedKCM (unsignedFourBitKCMCE clk ce) (tree (unsignedWeightedRegisteredAdder clk ce))

Note how much of the commonality of the basic structure of the combinational and pipelined KCMs has been captured by the unsignedKCM higher order circuit.

The layout produced for a pipelined unsigned 11-bit dynamic input and 11-bit constant is shown below for a Virtex part.

On a XCV300-4-BG432 part this circuit runs at a speed of at least 196MHz. The layout for an unsigned KCM with a 32-bit dynamic input and an 11-bit coefficient is shown below.

The 4-bit multiplications occur in the rectangular block on the left and the weighted adder tree is shown on the right hand side of the picture. On a XCV300-4-BG432 this circuit operates at a speed of at least 130MHz. Many other layouts are possible e.g. to place the 4-bit multiplications closer to their corresponding adders at the leaves of the adder tree.

###
**A Sorter Example in Lava**

### Sorting with Butterfly Networks

This section describes how to build a sorter circuit using butterfly networks which are carefully placed to ensure high performance. The sorter circuit is made by recursively merging the results of sub-sorts. A top-level schematic of the circuit that we present in this section is shown below. The merger that we present is bitonic which requires the first half of the input list to be increasing and the second half decreasing (or vice versa). The result of the top sorter is reversed to accommodate this requirement.

### Batcher's Bitonic Merger

Given the ability to sort two numbers and the diagram above we have a recursive formula for making sorters of any size. First the availability of a two sorter is assumed and the merger is designed. Then the implementation of the two sorter is given.

A merger called Batcher's Bitonic Merger can be made by using a butterfly of two sorters. Here is an example of a specific butterfly network of two sorters (written as 2S) which merges eight numbers:

To help describe such butterfly networks in Lava a few useful circuit combinators are introduced.

From the top level description we see that a reverse operation is required and we can simply use the built in Haskell reverse function:

Another very useful wiring combinator is called riffle and an instance of it is shown below:

This wiring combinator interleaves the odd and even elements of the input list (shown on the left). It can be defined in Lava as:

riffle = halve >-> ziP -> unpair

The halve function splits a list into two halves which are returned in a two element tuple. The ziP combinator takes a pair of lists and returns a new list of pairs by associating each element in the first list with the corresponding element in the second list. The unpair function then flattens this list of pairs into a list.

It is also useful to be able to perform the inverse function of riffle called unriffle. This circuit can be thought of as the reflection of the riffle circuit along a vertical axis as shown below.

The definition of unriffle in Lava is given below.

unriffle = pair >-> unzip >-> unhalve

Sometimes a bus containing

*n*elements is processed by using two copies of a circuit such that the first copy of the circuit operates on the bottom half of the input and the second copy of the circuit operates on the top half of the input as shown below for a four input bus:
The combinator that performs this task is called

**two**and is easily defined in Lava:**two**r = halve >->

**par**[r,r] >-> unhalve

Note that this combinator placed the first copy of r below the second copy of r. A specific instance of this combinator is

**two**two_sorter which is shown below:
Another combining form that uses two copies of the same circuit is

**ilv**(pronounced "interleave"). This combinator has the property that the bottom circuit processes the inputs at even positions and the top circuit processes the inputs at the odd positions. An instance of ilv R for an eight input bus is shown below.
The ilv combinator can be defined by noticing the it is the composition of an unriffle, two R and riffle:

ilv r = unriffle >=> two r >=> riffle

The evens combinator chops the input list into pairs and then applies copies of the same circuit to each input. The argument circuit for evens must be a pair to pair circuit. An instance of evens two_sorter over an eight input list is shown below.

This combinator is defined as:

evens f = chop 2 >-> maP f >-> concat

Using the combinators shown above we can now describe a butterfly network of some circuit r (such that r is a pair to pair circuit):

```
bfly r 1 = r
bfly r n = ilv (bfly r (n-1)) >-> evens r
```

Here is a picture of bfly r 1:

This makes sense in the case of a two sorter since a butterfly of size 1 has 2 inputs which can be sorted by a single two sorter. The layout for bfly r 2 is:

The left hand side of this picture shows an interleave of R and the right hand side shows evens R. The layout for bfly r 3 is:

Note that a sub-butterfly of size 2 has been identified with a pale background. It can be instructive to unfold the bfly r 3 expression and then try and spot where the various combinators occur in the picture.

bflyr 3 =ilv(bflyr 2)) >->evensr =ilv(ilvr >->evensr) >->evensr

To make a merger all we need to do is to instance this butterfly with a two sorter. Here is a picture of bfly r two_sorter (also shown before):

This solves the right hand side of the sorter architecture since bfly two_sorter makes a bitonic merger:

The two remaining sorters can be recursively decomposed using exactly the same technique used to decompose the top level sorter. For example, the upper sorter can be implemented by using a merger (shown on the right) and then sorting the two sub-lists. Since each sub-list contains just two elements we get to the base case of the recursion and deploy a two sorter.

But how is the merger realized? As before, it is just a butterfly of two sorters, in this case bfly 2 two_sorter:

Applying the same technique to the lower sorter gives the complete architecture for a size 3 sorter (i.e. 2^3 inputs = 8):

Although it is not at all obvious this circuit sorts eight numbers it has been systematically derived from a simple procedure which can be codified in Lava as:

sorter cmp 1 = cmp sorter cmp n = two (sorter cmp (n-1)) >-> sndList reverse >-> bfly cmp n

This description says that a sorter of degree 1 (i.e. 2 inputs) can be made using a two sorter. A larger sorter is made by using two small sorters, then reversing the result of the upper sort, and then merging these sub-sorts using a butterfly of two sorters. Note that this sorter description is parameterized on the specific sorter to be used.

An instance of sorter two_sorter produces a pipelined 8 input sorter which is shown below on a Virtex-II device:

Note that since Lava combinators include layout as well as connectivity information the resulting circuit occupies a rectangular area which corresponds to the pictures shown above.

A degree 4 (i.e. 16 inputs) sorter is shown below:

A degree 5 (i.e. 32 inputs) sorter is shown below:

A slightly larger version of the network above which sorts 32 16-bit numbers each clock tick operates at 165MHz on a XC2V3000 part with a latency of 14 clock ticks (using the 4.1i version of the Xilinx design tools). Larger data-sets can be sorted by storing numbers in BlockRAM and then iteratively sorting and merging. More details can be found in the paper Design and Verification of a Sorter Core.

Some instances of these butterfly networks were implemented on a XCV300 FPGA and ChipScope was used to verify that the real hardware actually did sort numbers:

#
**A 1D Systolic FIR**

### Introduction

This section describes how to build a one dimensional (1D) systolic filter in Lava. First we present the design of the processing element of the systolic array. Then we show how this processing element can be replicated to form a filter. Note that there are many ways of designing such filters and this technique is not the best possible implementation on Xilinx's FPGA architectures. This example has been chosen to help illustrate how to describe systolic-style systems in Lava.

**NOTE**: The filter implementations presented here are designed to illustrate aspects of Lava and are not the recommended implementations for Xilinx's FPGAs. Xilinx's Core Generator contains several optimized filter implementations.

### Finite Impulse Response Filters

The filter we shall build is called a finite impulse response filter (FIR) digital filter which calculate the weighted sum of the current and past inputs. FIR filters are also known as transversal filters or as a tapped delay line. The behavior of the finite impulse response filter can be described by the equation:

where yt denotes the output at time t and xt represents the input at time t and ak are the filter coefficients We shall use an "inner product" processing element to perform a single multiplication and addition and then replicate this processing element to make a circuit that compute the filtering function. We shall assume at the coefficients are constants which will allow us to use constant coefficient multipliers in our implementation.

We can model a finite impulse response filter in Haskell (the host language of Lava) with the following code.

semisystolicFIR weights xvals [] = [] semisystolicFIR weights xvals (xin:xrest) = sum [w * x | (w,x) <- zip weights xvals] : semisystolicFIR weights (xin : init xvals) xrest

This function takes three arguments:

- a list of weight values (weights)
- a list of the x-values at each tap of the filter (xvals)
- a list of input samples where the first sample value is called xin and the remainder are called xrest

If the input stream is empty i.e. the empty list [] then this function returns the empty list. If the input stream is non-empty then the first value in the input list is bound to xin and the remainder of the list is bound to xrest. The xvals are paired for multiplication up with their corresponding weight values by using the zip list processing function (for matching up elements pair-wise across two lists). The products are added to yield the filter result for this tick. This result is then followed by a list that corresponds to the remainder of the input stream (xrest) being recursively processed by the same function. However, we have to shift along the xvals so the xin value appears at the first tap. The expression (init xvals) returns the original xvals list with the last element removed. This allows us the express the "shifting in" of the new x value with the expression xin : init xvals.

Although the semisystolicFIR function does not correspond to a Lava circuit it can be used to simulate a filter. For example, assuming we have an input stream which corresponds to sine wave (scaled by 127) we can calculate the result of applying a finite impulse response filter with weights1 =[3, 9, 15, 7, 5]. Here is the results of performing such a simulation at the prompt of the ghci Haskell interpreter (looking at only the first 10 values of the input and output).

Systolic1DFir> take 10 sineValues [0,9,18,26,35,43,52,60,67,75] Systolic1DFir> take 10 (semisystolicFIR weights1 [0,0,0,0,0] sineValues) [0,0,27,135,375,672,1005,1340,1668,1997]

Systolic1DFir>

Systolic1DFir>

As a final step a filtering function may divide the output by the sum of the coefficients which will scale the output signal back into the range of the input signal (this is note done by this implementation). Choosing coefficients that have sum which is a power of two make it easy to perform such scaling since this corresponds to throwing allow some of the low order bits.

### A Semi-Systolic Filter

First we describe a semi-systolic filter. In a systolic design all the wires between processing elements have at least one latch and all the latches have the same clock signal. In a semi-systolic design this constraint is relaxed to allow wires between processing elements which do not have any latches.

The inner product processing element will take as inputs an accumulated sum from previous processing elements (yin), a filter coefficient (ai) and a sample value from the input stream (xin) and return two values: the xin is passed to xout and the yout is computed by performing the inner product calculation and adding it to the accumulated sum i.e. yout = yin + ai * xin. An implementation for an inner product processor is shown below where the purple circle denotes a constant coefficient multiplication by the value ai and the green circle denotes an adder. The input x values are passed from left to right and the accumulated sums as passed right right to left.

To make a complete filter we need a way of sequencing the x values one at a time through each processing element. This can be accomplished by placing a register at the xin input of the inner product processing giving the following processing element:

A semi-systolic filter can now be made by composing several comprises of this processing element. We need one processing element for each tap in the filter so a four tap filter would like look:

This is a semi-systolic filter because although there are registers on the wires carrying the x values there are no registers on the wires that carry the accumulated sum. The critical path of this circuit goes through four processing elements which makes this a poor implementation choice.

The Lava description of this filter can be used to generate VHDL for simulation and EDIF for implementation using Xilinx's place and route tools. The VHDL simulation shows that the filter behaves as expected:

The layout of this filter is shown below implemented in the corner of a Virtex-II FPGA (an XC2V1000-FF896 with speed grade 6).

This adders grow taller in a left to right direction as the accumulated sum gets larger. This implementation has a maximum combinational delay of 14.527ns (as reported by the place and route tools) which gives a maximum frequency of 68.8MHz and has 14 logic levels. This implementation uses 119 slices. This layout can be compacted by overlapping some of the pipeline registers with the result of the KCM calculation which also results in a faster implementation.

A better filter can be made by transforming this semi-systolic filter into a systolic filter by the systematic application of three techniques: retiming, slowdown and hold-up.

### The Processing Element

The processing element of the 1D systolic FIR is shown below. Both the x values and the accumulated results flow from left to right. Registers are added at the inputs and outputs for pipelining in a way that makes sure the accumulated sums and x values stay in synch.

This circuit can be specified by the Lava code below which makes use of a KCM and an adder that were presented earlier.

holdupPE clk k = fsT (registerAndMultiply clk k) >-> reorg >-> snD ((flexibleUnsignedAdder >|> vreg clk) >-> vreg clk) where reorg ((a,b),c) = (a,(b,c))

### A 4-tap Filter

An example of a four tap filter using this processing element is shown below:

This is formed by simply replicating the processing element horizontally. The x input has to be delayed by one clock tick to synchronize with the y inputs. When this filter is implemented by the Xilinx place and route tools the last two registers on the x path are automatically pruned as is the unused xout signal. The Lava code for this filter is shown below.

holdupFilter :: [Int] -> Bit -> [Bit] -> [Bit] holdupFilter weights clk xin = yout where (xout, yout) = hser [holdupPE clk k | k <- reverse weights]) (xin, replicate (length xin) gnd)

This filter has a much higher latency (8 ticks) than the semi-systolic filter:

However, since the longest combinational path between any two registers goes through just one processing element. The place and route tools report a maximum clock period of 6.687ns which allows this filter to be run at 150MHz.

A more compact but slower filter can be made by realizing the multiple register stages with SRL16 components (shift registers implemented in LUTs which can shift by up to 16 stages).

A better implementation can be produced by transforming the semi-systolic implementation into a systolic implementation by the systematic application of three techniques: retiming, slowdown and hold-up.

### Describing The Semi-Systolic Filter

Although the semi-systolic filter is not recommended for FPGA implementation we reproduce the implementation and show how it can be captured in Lava.

This architecture can not be directly described by the Lava combinators introduced so far because there is both left to right and right to left data-flow through each block. To help describe such communication patterns we introduce a new combinator called two-way serial and written as >< :

This combinator can then be used to describe a combinator called twoWayRow for the serial composition of many identical blocks that have two-way data-flow:

The definition of twoWayRow is:

twoWayRow = foldl1 (><)

This combinator can now be used directly to describe the semi-systolic filter shown above.