Built-in Chaining: Introducing Complex Components into Architectural Synthesis

Peter Marwedel, Birger Landwehr

Rainer D omer
Dept. of Information and Computer Science University of California, Irvine Irvine, CA 92697-3425, USA e-mail: [email protected]

Abstract In this paper, we extend the set of library components which are usually considered in architectural synthesis by components with built-in chaining. For such components, the result of some internally computed arithmetic function is made available as an argument to some other function through a local connection. These components can be used to implement chaining in a datapath in a single component. Components with built-in chaining are combinatorial circuits. They correspond to complex gates in logic synthesis. If compared to implementations with several components, components with built-in chaining usually provide a denser layout, reduced power consumption, and a shorter delay time. Multiplier/accumulators are the most prominent example of such components. Such components require new approaches for library mapping in architectural synthesis. In this paper, we describe an IP-based approach taken in our OSCAR synthesis system.

single component. In contrast, standard chaining requires external wiring between two or more components (therefore, we call it external chaining).

in1

in2

*
in3

Figure 1: Multiplier/adder In addition to multiplier/adders, multiplier/adder/accumulators (MACs), ALUs followed by (internal) shifters, and adders followed by (internal) comparators are components with internal chaining. The main advantages of such components include: The ability to generate efcient layout. The layout of BIC components is usually more efcient than the combined layout of several independently designed functional units. For example, abutment of corresponding lines may be possible. Delay and power consumption may be smaller than for separate components. The delay of BIC components may be small due to a) the more compact layout, b) exploitation of context-dependent information during logic synthesis, and c) adjustment of the strength of drivers. Design-reuse of complex components is facilitated. Exploiting the presence of BIC components is important for re-using available valuable designs. Example:
Table 1 shows area, delay times and power consumption of adders, multipliers, and multiplier-adders, respectively. The same information

I. MOTIVATION Architectural synthesis (also known as high-level synthesis (HLS)) can be dened as the task of implementing a given behavioural specication by means of an appropriate registertransfer (RT-) level architecture. Architectural synthesis is considered to provide the next productivity boost for designers of information-processing devices. RT-level components, which have been considered so far, include registers, register-les, busses, multiplexers and multifunctional units (ALUs). All functional units were considered to compute essentially a single (possibly control-selectable) standard function, such as addition, subtraction, or multiplication. On the other hand, current component libraries contain a growing number of components with built-in chaining (BIC) or internal chaining. With built-in chaining, the result of some computed standard function is made available as an argument to another function. For example, multiplier/adders multiply two numbers and add the result to a third one (see g. 1). With BIC components, chaining in HLS can be implemented by a
 This work has been supported by the Commission of the European Com-

munities under contract ESPRIT 6855 (LINK). In addition, the rst author was partially supported through NATO grant # CRG 950910.

Cell adder (64 Bit) multiplier (32 Bit) multiplier-adder-cell (32/64 bit) (built-in, internal chaining) multiplier-cells + adder-cells (external chaining) (32/64 bit)

Area [k2] 5.410 14.717 15.312 36.585

Power (@ 8 MHz) [mW] unknown 20.286 20.832 unknown

Propagation delay [ns] 18.44 107.76 108.65 117.66

Table 1: Area, power, and propagation delay for commercial library components

Figure 2: Layout for multiplier/adder with external vs. internal chaining

is also included for a multiplier-adder chain, built from individual multiplier and adder components (external chaining). In order to allow a fair comparison of the values, 64-bit adders have been used in all cases. The information has been generated with the help of the COMPASS DataPath Synthesizer for a 1  CMOS technology.

Fig. 2 shows the layout of the design with external chaining. Most importantly, the area for external chaining (computed as the bounding box of the combined layout), is more than two times larger than the BIC solution. This is caused by a very poor abutment of the two components. Some of the wiring area of the bounding box can possibly be saved, but external chaining will never be as efcient as internal chaining. The propagation delay has been computed with the COMPASS QTV timing analyser. Indicated values represent the maximum of the values for rising and falling edges. It is obvious, that the delay of the implementation with external chaining is less than the sum of the delays of the individual components (this is the reason for using non-additive delays models, see e.g. Rabaey et al.). Nevertheless, the delay is even smaller if internal chaining is used. BIC is important for many designs. Many digital signal processing architectures meet their constraints only because BIC is available. For example, many lter algorithms require MACs in order to meet their throughput constraints. Hence, the support of BIC in HLS is a must, despite the fact that the number of BIC components in libraries may not be very large.

The behavioural specication of digital signal processing architectures can be represented by data-ow graphs (DFGs). Fig. 3 shows the DFG of the well-known elliptical wave lter and indicates where multiplier/adders can be used. It is obvious that quite a number of operations can be implemented by multiplier/adders.

II. PREVIOUS WORK

In HLS synthesis, there has been only a very limited amount of work which took BIC components into account. Work on the Cathedral silicon compilers (see e.g. [7]) is one of the few contributions to the area. The execution unit (EXU) model of Cathedral includes BIC. However, only an essentially xed number of EXU types has been considered and no results are known which contribute to the topic of the paper. Research at the University of Eindhoven is directed at creating regular layouts from netlists in which regularity is not immediately obvious [8]. As a special case, these algorithms would be able to create a regular layout if, for example, an adder follows a multiplier in the netlist. However, the approach does not include exploitation of BIC components in HLS. In logic synthesis, the situation is different. So-called complex gates have been used in library mapping for many years (see [5] as an example).

in1 in2

in3
1 + + + + 2 5 6 7 + 2 * + + 2 * + + + + 2 12 13 14 * + + * + + 2 + * + + + * + + 2 + + * * + + 2 2 +

X +
out

2 3 4

8 9

* +
f

10 11

Figure 3: Use for multiplier/adders for elliptical wave lter

III. THE PROBLEM In order to support BIC components in HLS, several issues have to be considered. Lets start with an example to make these issues clear. Fig. 4 shows a section of a dataow-graph (DFG).

3. Implement + and  with a multiplier/adder in a single control step.

4. Implement + and  with a multiplier/adder performing a multi-cycle operation.

*
j
3

j j
2

Clearly, decisions have to take many factors into account: delay, cost (power and area), intended clocking frequency, predicted wiring delays, and surrounding operations. Also, one might want to consider testability and error-recovery aspects of the different design options [3].

In this context, it has to be mentioned, that well-established standard techniques in HLS have to be revisited:

Figure 4: DFG with associated operation labels 1. Scheduling and resource assignment can be modelled by functions taking DFG nodes (operations) as arguments and returning the corresponding control step and resource. In the context of BIC components, it is sometimes more adequate to use sets of nodes or operations as arguments.

Assume that a library containing adders, multipliers and multiplier/adders is given. There are several options for implementing this section of the DFG: 1. Implement + and  with two separate components in different control steps. 2. Implement + and  with two separate components in a single control step using standard chaining.

2. The function performed by a component can no longer be described by a single operation identier or an expression involving such a single identier. Rather, expressions including several operation identiers are required.

IV. APPROACH TAKEN IN OSCAR

A. Naming conventions
For our HLS system OSCAR (Optimum simultaneous scheduling, allocation and resource assignment), the ability to consider BIC components has been a major design goal. A very essential part of this is to select components for the nal architecture. We use binary decision variables for modelling the selection of certain components:

In OSCAR, complex operations in the DFG are labelled just like simple operations (see g. 4). Note that a certain node in the DFG may belong to several labels, to one for the node as a simple operation and possibly also to others if it is an element of complex operations. The essential task of HLS is to establish bindings between operation and component instances. Moreover, most HLS algorithms also generate bindings to control steps at which these operations are started. In OSCAR, control steps are represented by integers i from an index set I . Bindings are represented by decision variables x. In the case of predened instance bindings or cost functions containing interconnect costs, OSCAR uses triple-indexed decision variables with the following denition:

bk

1 0

; ;

if instance k is selected otherwise

(1)

Each k 2 K denotes a potential instance of a library element type m 2 M . Before synthesis is started, a sufcient set of instance identiers is made available for each m 2 M . Fig. 5 is a graphical representation of naming in our model. We distinguish between behavioural and structural domains. For each of these, we distinguish between types and instances. Specic types and instances are denoted by elements taken from sets of discrete elements. In actual implementations, sets of integers are used. In the paper, we are sometimes using sets of characters in order to improve readability. As can be seen from g. 5, the sets for operation instances, operation types, component instances and component types are denoted by J; G; K; and M , respectively. Component functionality is modelled by relation executable on. Def.: 8j 2 J; k 2 K : j executable on k  able to perform operation j 2 J .

xi;j;k =

8 1; : 0;

if operation j is started on component instance k at control step i otherwise

(2)

For each operation j 2 J , Rj  denotes the set of control steps during which j could be executed. Rj  can be computed by a simple ASAP/ALAP analysis or by techniques taking resource constraints into account.

k 2 K is
B. The model

We assume that all components are only able to start a limited number of operations. More precisely, we assume that component k is able to start a new operation after `j; k control steps if j is the operation most recently started. `j; k is called the component data initiation intervall (dii). Let G0 denote the set of standard operations which are supported for an HLS system. In the case of OSCAR, G0 corresponds to operations described in standard synthesis packages such as [9]. These operations can be represented by expressions containing single operation identiers, for example "+" and "". These operations are called simple operations. In addition to these operations, OSCAR considers complex operations or macro operations, operations which can be performed by BIC components. If, for example, the library contains a MAC, then ((in1  in2) + in3) will be considered as a complex operation. Let G00 denote complex operation types.

For a given component library, the cost function is a linear function of these variables.

X
m2M

COSTS 

m 

X
k2K typek=m

bk 

(3)

G = G0 G00. G denotes the set of all operation types. Operation instances corresponding to G; G0 and G00 are denoted by sets J; J 0 and J 00, respectively.
Def.:

This cost function is minimized under constraints describing correct solutions. The following constraints are not affected by the need of modelling BIC (see [6] for detailed equations):

layout

behaviour

J U J optype

G U G G

=
operation component

executable on

structure

type

executable on

instance

type

Figure 5: Naming conventions in the OSCAR model

1. Timing constraints: these constraints can be used to specify minimum and maximum delay between operations. Applied on read/writeoperations it allows to meet given timing specications. 2. Precedence constraints: These constraints reect that operations cannot be started before their arguments have been computed. 3. Chaining constraints These constraints reect the fact that the total combinatorial delay of the design should not exceed a given threshhold. In OSCAR, we assume that such a threshhold is known during the design process. An outer loop can be used to nd a good threshhold [4]. Using this threshhold, OSCAR generates constraints in case the combined delay of data-dependent operations exceeds the threshhold. Example:

Example:

Consider g. 6. For the sake of better readability, we use a; m; M to denote an adder, a multiplier and a MAC, respectively. Furthermore, we use +; ; to denote addition, multiplication and MAC operations.

If the macro operation is assigned to a MAC and to control step i = 2, then no other operation can be performed on this component for control steps in the range i;i + `j; k 1 = 2; 1 + `j;k  . This is the situation described in g. 6.

2
Since (5) does not distinguish between simple and complex operations, our model is able to handle components which can execute a mix of simple and complex operations. The form of equation (5) is very similar to resource constraints for other IP-based models. However, in our model, the set J includes complex operations. Also, two operations j1 ; j2 2 J may, in fact, represent overlapping segments of the DFG. In previous approaches [2], constant 1 is used on the right hand side of resource constraints. The current approach is required for HLS with integrated scheduling and assignment in order to avoid solutions, in which operations are assigned to non-selected component instances, i.e. instances for which bk is 0.

Assume that the combined delay of j1 and j2 in g. 4 exceeds the threshhold. Then, the following constraints can be generated:

8i 2 Rj1  Rj2  :
2

X
k

xi;j

;k + xi;j2 ;k   1

(4)

Chaining constraints are required in order to avoid solution 2 of section 3 in case high clocking frequencies have been specied. These constraints are not part of any other IPmodel we are aware of. Other constraints are affected by the need to model BIC: 4. Resource assignment constraints Resource assigment constraints guarantee that generated solutions respect the minimum data initiation intervall dii. diis are modelled by the following constraints:

5. Operation assignment constraints These constraints reect the fact that each operation has to be performed by a suitable hardware resource. This is guaranteed by the following constraints:

8k 2 K

X
j2J j executable on k

i+`X j;k ,1

xi ;j;k  bk
0

(5)

component instances

k=M (MAC)

k=a

k=m 1

*
x
x O xM 2O

=1

blocked for (j,k)= (O x,M) control steps

2 3 4 5 6
control steps

Figure 6: Graphical interpretation of assignment constraints

X X xi;j ;k + 8j 0 2 J 0 : k K i2Rj  j exec on k

V. RESULTS In the following, we will describe our results for two standard examples: the elliptical wave lter (EWF) and an edgedetection algorithm. All results use the components described in table 1. The cycle time has been set to 50 ns in order to allow a controller and multiplexer delay of about 34 ns for single-cycle adds. For this cycle time, multiplications and multiply/adds will always be multi-cycle operations. None of the results uses external chaining. Hence, the following tables mainly demonstrate the effect of built-in chaining. The execution times have been measured on a SPARC 20 using the mixed IP-solver lp solve [1]. Tables 2 and 3 show the results for the elliptical wave lter. add cs 20 21 22 3 2 2 4 3 2 mult active area [k2] 66.083 48.961 34.244 runtime [s]

00 : 2K j0002J00 i2Rj 00  j00kexec on k j 2j

xi;j ;k = 1
00

(6)

The meaning of these constraints is as follows: Each simple operation j 0 has to be implemented. There are two ways of doing this: a) j 0 is implemented individualy. In this case, the sum of xi;j ;k over all possible control steps Rj 0  and all
0

components will be 1 (1st term in (6) = 1). b) j 0 is implemented as part of a macro operation j 00 . j 0 may actually be part of several macro operations, but only one of these can be implemented as a macro operation. Hence, the sum of xi;j 00;k over all macro operations j 00 enclosing j 0 , over all control steps Rj 00, and over all resources k capable of executing j 00 must be one (second term in (6)). Example:
Let us assume that R+ = 3; R  = 2; R  = 2::3 . Then, the following equations will be generated:

1 1 3

Table 2: Results for EWF without BIC components

j0 =  j0 = +
2

: :

x2m + x2 x3+a + x2

M + x3 M = 1 M + x3 M = 1

(7) (8)

add cs 17 18 19 3 2 2

mult

mac

These equations guarantee that either the macro operation or the simple operations will be bound to a hardware component.

1 1 0

3 2 2

[k2 ] 67.872 50.154 35.434

active area

runtime [s]

1 37 216

By considering simple operations as a special case of macro operations, it would be possible to use just the right term of equation 6, but this would lead to some redundant computations.

Table 3: Results for EWF with BIC components Note that the use of BIC can be exploited in two ways: a) to

get a smaller design if the cycle budget is xed and b) to get a faster design if the area budget is xed. Tables 4 and 5 show the results for the edge detector algorithm. add cs 12 13 14 2 2 2 2 2 2 8 7 4 sub mult active area runtime [s]

These results demonstrate the efciency of designs using BIC. The authors appreciate the comments of Fadi Kurdahi and Nikil Dutt (UC Irvine) on an earlier version of the manuscript.

[k2 ] 127.418 112.700 68.548

REFERENCES [1] M.R.C.M. Berkelaar. Unixtm manual page of lp solve. Eindhoven University of Technology, Design Automation Section, 1992. [2] C. H. Gebotys and M. I. Elmasry. Optimal VLSI Architectural Synthesis. Kluwer Academic Publishers, 1992. [3] I. G. Harris and A. Orailoglu. Microarchitectural synthesis of VLSI designs with high test concurrency. 31st ACM/IEEE Design Automation Conference, pages 206 211, 1994. [4] P. K. Jha, S. Parameswaran, and N. D. Dutt. Reclocking controllers for minimum execution time. Technical Report 94-40, Information and Computer Science, University of California at Irvine, 1994. [5] K. Keutzer. DAGON: Technology binding and local optimization by DAG matching. 24th Design Automation Conference, pages 341347, 1987. [6] B. Landwehr, P. Marwedel, and R. D omer. OSCAR: Optimum simultaneous scheduling, allocation and resource binding based on integer programming. Euro-DAC, 1994. [7] H. De Man, J. Rabaey, and P. Six. CATHEDRAL II: A synthesis and module generation system for multiprocessor systems on a chip. in: G.DeMicheli, A.SangiovanniVincentelli, P.Antognetti: Design Systems for VLSI CircuitsLogic Synthesis and Silicon Compilation, Martinus Nijhoff Publishers, 1987. [8] R. Nijssen and J.A.G. Jess. Data path regularity extraction. IFIP Workshop on Logic and Architecture Synthesis, 1994. [9] Special Interest Group on Synthesis from VHDL. VHDL arithmetic package for synthesis. repository at INTERNET host vhdl.org, login anonymous, le vi/vhdlsynth/numeric bit.vhd, 1993.

1 1 57

Table 4: Results for edge detector without BIC components

add cs 12 13 14 2 1 1

sub

mult

mac

2 2 1

4 3 3

1 1 1

active area [k2 ] 83.860 66.737 64.304

runtime [s] 2 10 192

Table 5: Results for edge detector with BIC components

VI. CONCLUSION In this paper, we have proposed to pay attention to the need of modelling components with built-in chaining (BIC). We have stressed the importance of the support of these components in high-level synthesis. As an example, we have shown how BIC can be modelled in integer-programming (IP) based synthesis algorithms. IP-models provide a basis for adding support for BIC in a rather straightforward way. Due to recent advances in IP-based modelling, we have been able to generate designs in acceptable computation time [6].