Design apparatus and a method for generating an implementable description of a digital system

[0001] This application is a continuation of and claims priority to and incorporates by reference, in its entirety, U.S. application Ser. No. 09/237,549, titled “A DESIGN APPARATUS AND A METHOD FOR GENERATING AN IMPLEMENTABLE DESCRIPTION OF A DIGITAL SYSTEM”, filed Jan. 26, 1999, now U.S. Pat. No. 6,606,588, which in turn is a continuation of and claims priority to U.S. application Ser. No. 09/041,985, titled “DESIGN APPARATUS AND A METHOD FOR GENERATING AN IMPLEMENTABLE DESCRIPTION OF A DIGITAL SYSTEM”, filed Mar. 13, 1998, now U.S. Pat. No. 6,233,540, which in turn claims priority under 35 U.S.C. § 120, to the following U.S. provisional patent applications: “Design Environment and a Method for Dataflow Support and Refinement of Dataflow for Hardware Design and Hardware/software Co-design,” Application No. 60/039,078, and filed on Mar. 14, 1997; “Design Environment and a Method for Generating an Implementable Description of a Digital System,” Application No. 60/039,079, and filed on Mar. 14, 1997; “Design Environment and a Method for Generating an Implementable Description of a Digital System,” Application No. 60/041,121, and filed on Mar. 20, 1997”.

[0002] The present invention is situated in the field of design of systems. More specifically, the present invention is related to a design apparatus for digital systems, generating implementable descriptions of said systems.

[0003] The present invention is also related to a method for generating implementable descriptions of said systems.

[0004] The current need for digital systems forces contemporary system designers with ever increasing design complexities in most applications where dedicated processors and other digital hardware are used, demand for new systems is rising and development time is shortening. As an example, currently there is a high interest in digital communication equipment for public access networks. Examples are modems for Asymmetric Digital Subscriber Loop (ADSL) applications, and up- and downstream Hybrid Fiber-Coax (HFC) communication. These modems are preferably implemented in all-digital hardware using digital signal processing (DSP) techniques. This is because of the complexity of the data processing that they require. Besides this, these systems also need short development cycles. This calls for a design methodology that starts at high level and that provides for design automation as much as possible.

[0005] One frequently used modeling description language is VHDL (VHSIC Hardware Description Language), which has been accepted as an IEEE standard since 1987. VHDL is a programming environment that produces a description of a piece of hardware. Additions to standard VHDL can be to implement features of Object Oriented Programming Languages into VHDL. This was described in the paper OO-VHDL (Computer, October 1995, pages 18-26). Another frequently used modeling description language is VERILOG.

[0006] A number of commercially available system environments support the design of complex DSP systems.

[0007] MATLAB of Mathworks Inc offers the possibility of exploration at the algorithmic level. It uses the data-vector as the basic semantical feature. However, the developed MATLAB description has no relationship to a digital hardware implementation, nor does MATLAB support the synthesis of digital circuits.

[0008] SPW of Alta Group offers a toolkit for the simulation of these kind of systems. SPW is typically used to simulate data-flow semantics. Data-flow semantics define explicit algorithmic iteration, whereas data-vector semantics do not. SPW relies on an extensive library and toolkit to develop systems. Unlike MATLAB, the initial description is a block-based description. Each block used in the systems appears in two different formats, (a simulatable and a synthesizable version) which results in possible inconsistency.

[0009] COSSAP of Synopsys performs the same kind of system exploration as SPW.

[0010] DC and BC are products of Synopsys that support system synthesis. These products do not provide sufficient algorithm exploration functions.

[0011] Because all of these tools support only part of the desired functionality, contemporary digital systems are designed typically with a mix of these environments. For example, a designer might do algorithmic exploration in MATLAB, then do architecture definition with SPW, and finally map the architecture definition to an implementation in DC.

[0012] It is an aim of the present invention to disclose a design apparatus that allows to generate from a behavioral description of a digital system, an implementable description for said system.

[0013] It is another aim of the present invention to disclose a the design apparatus that allows for design, digital systems starting from a data vector or data flow description and generating an implementable level such as VHDL. A further aim is to perform such design tasks within one object oriented environment.

[0014] Another aim is to provide a means comprised in said design apparatus for simulating the behavior of the system at any level of the design stage or trajectory.

[0015] A first aspect of the present invention concerns a design apparatus compiled on a computer environment for generating from a behavioral description of a system comprising at least one digital system part, an implementable description for said system, said behavioral description being represented on said computer environment as a first set of objects with a first set of relations therebetween, said implementable description being represented on said computer environment as a second set of objects with a second set of relations therebetween, said first and second set of objects being part of a design environment.

[0016] A behavioral description is a description which substantiates the desired behavior of a system in a formal way. In general, a behavioral description is not readily implementable since it is a high-level description, and it only describes an abstract version of the system that can be simulated. An implementable description is a more concrete description that is, in contrast to a behavioral description, detailed enough to be implemented in software to provide an approximative simulation of real-life behavior or in hardware to provide a working semiconductor circuit.

[0017] With design environment is meant an environment in which algorithms can be produced and run by interpretion or compilation.

[0018] With objects is meant a data structure which shows all the characteristics of an object from an object oriented programming language, such as described in “Object Oriented Design” (G. Booch, Benjamin/Cummings Publishing, Redwood City, Calif., 1991).

[0019] Said first and second set of objects are preferably part of a single design environment.

[0020] Said design environment comprises preferably an Object Oriented Programming Language (OOPL). Said OOPL can be C++.

[0021] Said design environment is preferably an open environment wherein new objects can be created. A closed environment will not provide the flexibility that can be obtained with an open environment and will limit the possibilities of the user.

[0022] Preferably, at least part of the input signals and output signals of said first set of objects are at least part of the input signals and output signals of said second set of objects. Essentially all of the input signals and output signals of said first set of objects can be essentially all of the input signals and output signals of said second set of objects.

[0023] At least part of the input signals and output signals of said behavioral description are preferably at least part of the input signals and output signals of said implementable description. Essentially all of the input signals and output signals of said behavioral description can be essentially all of the input signals and output signals of said implementable description.

[0024] Said first set of objects has preferably first semantics and said second set of objects has preferably second semantics. With semantics is meant the model of computation. Said first semantics is preferably a data-vector model and/or a data-flow model. Said second semantics is preferably a Finite State Machine Data Path (FSMD) data structure, comprising a control part and a data processing part, the data processing part being modeled by a signal flow graph (SFG) data structure and the control part being modeled by a FSM data structure. The terms FSMD and SFr are used interchangeably throughout the text.

[0025] Preferably, the impact in said implementable description of at least a part of the objects of said second set of objects is essentially the same as the impact in said behavioral description of at least a part of the objects of said first set of objects.

[0026] Preferably, the impact in said implementable description of essentially all of the objects of said second set of objects is essentially the same as the impact in said behavioral description of essentially all of the objects of said first set of objects.

[0027] With impact is meant not only the function, but also the way the object interacts with its environment from an external point of view. A way of rephrasing this is that the same interface for providing input and collecting output is present. This does not mean that the actual implementation of the data-processing between input and output is the same. The implementation is embodied by objects, which can be completely different but perform a same function. In an OOPL, the use of methods of an object without knowing its actual implementation is referred to as information hiding.

[0028] The design apparatus preferably further comprises means for simulating the behavior of said system said means simulating the behavior of said behavioral description, said implementable description or any intermediate description therebetween. Said intermediate description can be obtained after one or several refining steps from said behavioral description.

[0029] Preferably, at least part of said second set of objects is derived from objects belonging to said first set of objects. This can be done by using the inheritance functionalities provided in an OOPL. Essentially all of said second set of objects can be derived from objects belonging to said first set of objects.

[0030] Said implementable description can be at least partly obtained by refining said behavioral description. Said implementable description can be essentially obtained by refining said behavioral description. Preferably, said refining comprises the refining of objects.

[0031] The design apparatus can further comprise means to derive said first set of objects from a vector description, preferably a MATLAB description, describing said system as a set of operations on data vectors, means for simulating statically or demand-driven scheduled dataflow on said dataflow description and/or means for clock-cycle true simulating said digital system using said dataflow description and/or one or more of said SFG data structures.

[0032] In a preferred embodiment, said implementable description is an architecture description of said system, said system advantageously further comprising means for translating said architecture description into a synthesizable description of said system, said synthesizable description being directly implementable in hardware. Said synthesizable description is preferably a netlist of hardware building blocks. Said hardware is preferably a semiconductor chip or a electronic circuit comprising semiconductor chips.

[0033] A synthesizable description is a description of the architecture of a semiconductor that can be synthesized without further processing of the description. An example is a VHDL description.

[0034] Said means for translating said architecture description into a synthesizable description can be Cathedral-3 or Synopsys DC.

[0035] A second aspect of the present invention is a method for designing a system comprising at least one digital part, comprising a refining step wherein a behavioral description of said system is transformed into an implementable description of said system, said behavioral description being represented as a first set of objects with a first set of relations therebetween and said implementable description being represented as a second set of objects with a second set of relations therebetween.

[0036] Said refining step preferably comprises translating behavioral characteristics at least partly into structural characteristics. Said refining step can comprise translating behavioral characteristics completely into structural characteristics.

[0037] Said method can further comprise a simulation step in which the behavior of said behavioral description, said implementable description and/or any intermediate description therebetween is simulated.

[0038] Said refining step can comprises the addition of new objects, permitting interaction with existing objects, and adjustments to said existing objects allowing said interaction.

[0039] Preferably, said refining step is performed in an open environment and comprises expansion of existing objects. Expansion of existing objects can include the addition to an object of methods that create new objects. Said object is said to be expanded with the new objects. The use of expandable objects allows to use meta-code generation: creating expandable objects implies an indirect creation of the new objects.

[0040] Said behavioral description and said implementable description are preferably represented in a single design environment, said single design environment advantageously being an Object Oriented Programming Language, preferably C++.

[0041] Preferably, said first set of objects has first semantics and said second set of objects has second semantics. Said first semantics is preferably a data-vector model and/or a data-flow model. Said second semantics is preferably an SFG data structure.

[0042] The refining step comprises preferably a first refining step wherein said behavioral description being a data-vector model is at least partly transformed into a data-flow model. Advantageously, said data-flow model is an untimed floating point data-flow model.

[0043] Said refining step preferably further comprises a second refining step wherein said data-flow model is at least partly transformed into an SFG model. Said data-flow model can be completely transformed into an SFG model.

[0044] In a preferred embodiment, said first refining step comprises the steps of determining the input vector lengths of input, output and intermediate signals, determining the amount of parallelism of operations that process input signals under the form of a vector to output signals, determination of objects, connections between objects and signals between objects of said data-flow model, and determining the wordlength of said signals between objects. In the sequel of this application, the term “actors” is also used to denote objects. Connections between objects are denoted as “edges” and signals between objects are denoted as “tokens”. Said step of determining the amount of parallelism can preferably comprise determining the amount of parallelism for every data vector and reducing the unspecified communication bandwidth of said data-vector model to a fixed number of communication buses in said data-flow model. Said step of determination of actors, edges and tokens of said data-flow model preferably comprises defining one or a group of data vectors in said first data-vector model as actors; defining data precedences crossing actor bounds, as edges, said edges behaving like queues and transporting tokens between actors; construct a system schedule and run a simulation on a computer environment. Said second refining step comprises preferably transforming said tokens from floating point to fixed point. Preferably, said SFG model is a timed fixed point SFG model.

[0045] Said second set of objects with said second set of relations therebetween are preferably at least partly derived from said first set of objects with said first set of relations therebetween. Objects belonging to said second set of objects are preferably new objects, identical with and/or derived by inheritance from objects from said first set of objects, or a combination thereof.

[0046] Several of said SFG models can be combined with a finite state machine description resulting in an implementable description. Said implementable description can be transformed to synthesizable code, said synthesizable code preferably being VHDL code.

[0047] Another aspect of the present invention is a method for simulating a system, wherein a description of a system is transformed into compilable C++ code.

[0048] Preferably, said description is an SFG data structure and said compilable C++ code is used to perform clock cycle true simulations.

[0049] Several SFG data structures can be combined with a finite state machine description resulting in an implementable description, said implementable description being said compilable C++ code suitable for simulating said system as software.

[0050] A clock-cycle true simulation of a system uses one or more SFG data structures.

[0051] Said clock-cycle true simulation can be an expectation-based simulation, said expectation-based simulation comprising the steps of: annotating a token age to every token; annotating a queue age to every queue; increasing token age according to the token aging rules and with the travel delay for every queue that has transported the token; increasing queue age with the iteration time of the actor steering the queue, and; checking whether token age is never smaller than queue age throughout the simulation.

[0052] Another aspect of the present invention is a hardware circuit or a software simulation of a hardware circuit designed with the design apparatus as recited higher.

[0053] Another aspect of the present invention is a hardware circuit or a software simulation of a hardware circuit designed with the method as recited higher.

[0054] The present invention will be further explained by means of examples, which does not limit the scope of the invention as claimed.

[0079] The present invention can be described as a design environment for performing subsequent gradual refinement of descriptions of digital systems within one and the same object oriented programming language environment. The lowest level is semantically equivalent to a behavioral description at the register transfer (RT) level.

[0080] A preferred embodiment of the invention comprising the design method according to the invention is called OCAPI. OCAPI is part of a global design methodology concept SOC++. OCAPI includes both a design environment in an object oriented programming language and a design method. OCAPI differentiates from current systems that support architecture definition (SPW, COSSAP) in the way that a designer is guided from the MATLAB level to the register transfer level. This way, combined semantic and syntactic translations in the design flow are avoided.

• • [0081] The designer is offered a single coding framework in an object oriented programming language, such as C++, to express refinements to the behavior. An open environment is used, rather than the usual interface-and-module approach.
  • [0082] The coding framework is a container of design concepts, used in traditional design practice. Some example design concepts currently supported are simulation queues, finite state machines, signal flowgraphs, hybrid floating/fixed point data types, operation profiling and signal range statistics. The concepts take the form of object oriented programming language objects (referred to as object in the remainder of this text), that can be instantiated and related to each other.
  • [0083] With this set of objects, a gradual refinement design route is offered: more abstract design concepts can be replaced with more detailed ones in a gradual way. Also, design concepts are combined in an orthogonal way: quantization effects and clock cycles (operation/operator mapping) for instance are two architecture features that can be investigated separately. Next, the different design hierarchies can be freely intermixed because of this object-oriented approach. For instance, it is possible to simulate half of the description at fixed point level, while the other half is still in floating point.
  • [0084] The use of a single object oriented programming language framework in OCAPI allows fast design iteration, which is not possible in the typical nowadays hybrid approach.

[0085] Comparing to existing data-flow-based systems like SPW and COSSAP we see that the algorithm iterations can be freely chosen. Comparing to existing hardware design environments like DC or BC, we see that we can start from a specification level that is more abstract than the connection of blocks.

[0086] Two concepts of scaleable parallelism and expectation based simulation are introduced. The designer is given an environment to check the feasibility of what the designer thinks that can be done. In the development process, the designer creates his library of Signal FlowGraph (SFG) versions of abstract MATLAB operations.

[0087] OCAPI is a C++ library intended for the design of digital systems. It provides a short path from a system design description to implementation in hardware. The library is suited for a variety of design tasks, including:

• • [0088] Fixed Point Simulations
  • [0089] System Performance Estimation
  • [0090] System Profiling
  • [0091] Algorithm-to-Architecture Mapping
  • [0092] System Design according to a Dataflow Paradigm
  • [0093] Verification and Testbench Development
    Development Flow

[0094] The Flow Layout

[0095] The design flow according to an embodiment of the present invention, as shown in FIG. 1D, starts off with an untimed, floating point C++ system description 101. Since data-processing intensive applications such as all-digital transceivers are targeted, this description uses data-flow semantics. The system is described as a network of communicating components.

[0096] At first, the design is refined, and in each component, features expressing hardware implementation are introduced, including time (clock cycles) and bittrue rounding effects. The use of C++ allows to express this in an elegant way. Also, all refinement is done in a single environment, which greatly speedups the design effort.

[0097] Next, the timed, bittrue C++ description 103 is translated into an equivalent HDL description by code generation. For each component, a controller description 105 and a datapath description 107 can be generated. Also for each component a single HDL description can be generated, this description preferably jointly representing the control processing and data processing of the component. This is done because OCAPI relies on separate synthesis tools for both parts, each one optimized towards controller or else datapath synthesis tasks. Through the use of an appropriate object modeling hierarchy the generation of datapath and controller HDL can be done fully automatic.

[0098] For datapath synthesis 109, OCAPI relies on the Cathedral-3 datapath synthesis tools, that allow to obtain a bitparallel hardware implementation starting from a set of signal flowgraphs. Controller synthesis 111 on the other hand is done by the logic synthesis of Synopsys DC. This divide and conquer strategy towards synthesis allows each tool to be applied at the right place.

[0099] During system simulation, the system stimuli 113 are also translated into testbenches that allow to verify the synthesis result of each component. After interconnecting all synthesized components into the system netlist, the final implementation can also be verified using a generated system testbench 115.

[0100] The System Model

[0101] The system machine model that is used is a set of concurrent processes. Each process translates to one component in the final system implementation.

[0102] At the system level, processes execute using data flow simulation semantics. That is, a process is described as an iterative behavior, where inputs are read in at the start of an iteration, and outputs are produced at the end. Process execution can start as soon as the required input values are available.

[0103] Inside of each process, two types of description are possible. The first one is an untimed description, and can be expressed using any C++ constructs available. A firing rule is also added to allow dataflow simulation. Untimed processes are not subject to hardware implementation but are needed to express the overal system behavior. A typical example is a channel model used to simulate a digital transceiver.

[0104] The second flavor of processes is timed. These processes operate synchronously to the system clock. One iteration of such a process corresponds to one clock cycle of processing. Such a process falls apart in two pieces: a control description and a data processing description.

[0105] The control description is done by means of a finite state machine, while the data description is a set of instructions. Each instruction consists of a series of signal assignments, and can also define process in- and outputs. Upon execution, the control description is evaluated to select one or more instructions for execution. Next, the selected instructions are executed. Each instruction thus corresponds to one clock cycle of RT behavior.

[0106] For system simulation, two schedulers are available. A dataflow scheduler is used to simulate a system that contains only untimed blocks. This scheduler repeatedly checks process firing rules, selecting processes for execution as their inputs are available. When the system also contains timed blocks however, a cycle scheduler is used. The cycle scheduler manages to interleave execution of multi-cycle descriptions, but can incorporate untimed blocks as well.

[0107] The Standard Program

[0108] The library of OCAPI has been developed with the g++ C++ GNU compiler. The best mode embodiment uses the g++ 2.8.1 compiler, and has been successfully compiled and run under the HPUX 10 (HPUX10) operating system platform. It is also possible to use a g++ 2.7.2 compiler, allowing for compilation and run under operating system platforms such as HPUX-9 (HPRISC), HPUX-10 (HPUX10), SunOS (SUN4), Solaris (SUN5) and Linux 2.0.0 (LINUX).

[0109] The layout of the ‘standard’ g++ OCAPI program will be explained, including compilation and linking of this program.

[0110] First of all, g++ is a preferred standard compilation environment. On Linux, this is already the case after installation. Other operating system vendors however usually have their own proprietary C++ compiler. In such cases, the g++ compiler should be installed on the operating system, and the PATH variable adapted such that the shell can access the compiler.

[0111] The OCAPI library comes as a set of include files and a binary lib. All of these are put into one directory, which is called the BASE directory.

[0112] The ‘standard program’ is the minimal contents of an OCAPI program. It has the following layout.

[0113]

	include ‘‘qlib.h’’
	int main ( )
	{
	//your program goes here
	}

[0114] The include “qlib.h” includes everything you need to access all classes within OCAPI.

[0115] If this program is called “standard.cxx”, then the following makefile will transform the source code into an executable for you:

[0116]

	HOSTTYPE = HPUX10
	BASE = /imec/vsdm/OCAPI/release/v0.9
	CC = g++
	QFLAGS = -c -g -Wall -I${BASE}
	LIBS = -lm
	%.o: %.cxx
	$(CC) $(QFLAGS) $< -o $@
	TARGET = standard
	all: $(TARGET)
	define lnkqlib
	$(CC) ${circumflex over ( )} -o $@ $(LIBS)
	endef
	OBJS = standard.o
	standard:${OBJS} $(BASE)/lib$(HOSTTYPE) qlib.a
	${lnkqlib}
	clean:
	rm -f *.o $(TARGET)

[0117] This is a makefile for GNU's “make”; other “make” programs can have a slightly different syntax, especially for the definition of the “lnkqlib” macro. It is not the shortest possible solution for a makefile, but it is one that works on different platforms without making assumptions about standard compilation rules.

[0118] The compilation flags “QFLAGS” mean the following: “-c” selects compilation-only, “-g” turns on debugging information, and “-Wall” is the warning flag. The debugging flag allows you to debug your program with “gdb”, the GNU debugger.

[0119] Even if you don't like a debugger and prefer “printf( )” debugging, “gdb” can at least be of great help in the case the program core dumps. Start the program under “gdb” (type “gdb standard” at the shell prompt), type “run” to let “standard” crash again, and then type “bt”. One now see the call trace.

Calculation

[0120] OCAPI processes both floating point and fixed point values. In contrast to the standard C++ data types like “int” and “double”, a “hybrid” data type class is used, that simulates both fixed point and floating point behavior.

[0121] The dfix Class

[0122] This class is called “dfix”. The particular floating/fixed point behavior is selected by the class constructor. The standard format of this constructor is

[0123]

	dfix a;    // a floating point value
	dfix a (0.5);// a floating point value with initial value
	dfix a (0.5, 10, 8);
	// a fixed point value with initial value,
	// 10 bits total word-length, 8 fractional bits

[0124] A fixed point value has a maximal precision of the mantissa precision of a C++ “double”. On most machines, this is 53 bits.

[0125] A fixed point value can also select a representation, an overflow behavior, and a rounding behavior. These flags are, in this order, optional parameters to the “dfix” constructor. They can have the following values.

• • [0126] Representation flag: “dfix::tc” for two's complement signed representation, “dfix::ns” for unsigned representation.
  • [0127] Overflow flag: “dfix::wp” for wrap-around overflow, “dfix::st” for saturation.
  • [0128] Rounding flag: “dfix::fl” for truncation (floor), “dfix::rd” forrounding behavior.

[0129] Some examples are

[0130]

	dfix a(0.5, 10, 8);
	// the default is two's complement, wrap-around,
	// truncated quantisation
	dfix a(0.5, 10, 8, dfix::tc, dfix::st, dfix::rd);
	// two's complement, saturation, rounding quantisation
	dfix a(0.5, 10, 8, dfix::ns);
	// unsigned, wrap-around, truncated quantisation

[0131] When working with fixed point “dfix”es, it is important to keep the following rule in mind: “quantisation occurs only when a value is defined or assigned”. This means that a large expression with several intermediate results will never have these intermediate values quantised. Especially when writing code for hardware implementation, this should be kept in mind. Also intermediate results are stored in finite hardware and therefore will have some quantisation behavior. There is however a a “cast” operator that will come at help here.

[0132] The dfix Operators

[0133] The operators on “dfix” are shown below

• • [0134] +, −, *, /
    • [0135] Standard addition, subtraction (including unary minus), multiplication and division.
  • [0136] +=, −=, *=, /=
    • [0137] In-place versions of previous operators.
  • [0138] abs
    • [0139] Absolute value.
  • [0140] <<, >>
    • [0141] Left and right shifts.
  • [0142] <<=, >>=
    • [0143] In place left and right shifts.
  • [0144] msbpos
    • [0145] Most-significant bit position.
  • [0146] &, |, ^, ˜
    • [0147] Bitwise and, or, exor, and not operators.
  • [0148] frac( ) (member call)
    • [0149] Fractional part.
  • [0150] ==, !=, <=, >=, <, >
    • [0151] Relational operators: equal, different, smaller then or equal to, greater then or equal to, smaller then, greater then. These return an “int” instead of a “dfix”.

[0152] All operators with exception of the bitwise operators work on the maximal fixed point precision (53 points). The bitwise operators have a precision of 32 bits (a C++ “long”). Also, they assume the fixed point representation contains no fractional bits.

[0153] In addition to the arithmetic operators, several utility methods are available for the “dfix” class.

[0154]

	dfix a,b;
	// cast a to another type
	b = cast(dfix(0, 12, 10), a);
	// assign b to a, retaining the quantisation of a
	a = b;
	// assign b to a, including the quantisation
	a.duplicate(b);
	// return the integer part of b
	int c = (int) b;
	// retrieve the value of b as a double
	double d,e:
	d = b.Val( );
	e = Val (b);
	// return quantisation characteristics of a
	a.TypeW( );	// returns the number of bits
	a.TypeL( );	// returns the number of fractional bits
	a.TypeSign( );	// returns dfix::tc or dfix::ns
	a.TypeOverflow( );	// returns dfix::wp or dfix::st
	a.TypeRound( );	// returns dfix::fl or dfix::rd
	// check if two dfixes are identical in value and
	quantisation
	identical(a,b);
	// see wether a is floating or fixed point
	a.TypeMode( ); // returns dfix::fixpoint or dfix::floatpoint
	a.isDouble( );
	a.isFix( );
	// write a to cout
	cout << a;
	// write a to stdout, in float format,
	// on a field of 10 characters
	write (cout, a, ‘f’, 10);
	// now use a fixed-format
	write (cout, a, ‘g’, 10);
	// next assume a is a fixed point number, and write out an
	// integer representation (considering the decimal point at
	// the lsb of a) use a hexadecimal format
	write (cout, a, ‘x’, 10);
	// use a binary format
	write (cout, a, ‘b’, 10);
	// use a decimal format
	write (cout, a, ‘d’, 10);
	// read a from stdin
	cin >> a;

Communication

[0155] Apart from values, OCAPI is concerned with the communication of values in between blocks of behavior. The high level method of communication in OCAPI is a FIFO queue, of type “dfbfix”. This queue is conceptually infinite in length. In practice it is bounded by a sysop phonecall telling that you have wasted up all the swap space of the system.

[0156] The dfbfix Class

[0157] A queue is declared as

[0158] dfbfix a(“a”);

[0159] This creates a queue with name a. The queue is intented to pass value objects of the type “dfix”. There is also an alias type of “dfbfix”, known as “FB” (flow buffer). So you can also write

• • [0160] FB a(“a”);

[0161] The dfbfix Operations

[0162] The basic operations on a queue allow to store and retrieve “dfix” objects. The operations are

[0163]

	dfix k;
	dfix j(0.5);
	dfbfix a(‘‘a’’);
	// insert j at the front of a
	a.put(j);
	// operator format for an insert
	a << j;
	// insert j at position 5, with position 0 corresponding to
	// the front of a.
	a.putIndex(j,5);
	// read one element from the back of a
	k = a.get( );
	// operator format for a read
	a >> j;
	// peek one element at position 1 of a
	k = a.getIndex(1);
	// operator format for peek
	k = a[1];
	// retrieve one element from a and throw it
	a.pop( );
	// throw all elements, if any, from a
	a.clear ( );
	// return the number of elements in a as an int
	int n = a.getSize( );
	// return the name of the queue
	char *p = a.name( );

[0164] Whenever you perform an access operation that reads past the end of a FIFO, a runtime error results, showing

• [0165] Queue Underflow @ get in queue a

[0166] Utility Calls for dfbfix

[0167] Besides the basic operations on queues, there are some additional utiliy operations that modify a queue behavior

[0168]

	// make a queue of length 20. The default length of a queue
	// is 16. Whenever this length is exceeded by a put, the
	// storage in the queue is dynamically expanded by a factor
	// of 2.
	dfbfix a(‘‘a’’, 20);
	// After the asType( ) call, the queue will have an input
	// ‘‘quantizer’’ that will quantize each element inserted
	// into the queue to that of the quantizer type
	dfix q(0, 10, 8);
	a.asType(q);
	// After an asDebug( ) call, the queue is associated with a
	// file, that will collect every value written into the
	// queue. The file is opened as the queue is initialized
	// and closed when the queue object is destroyed.
	a.asDebug (‘‘thisfile.dat’’);
	// Next makes a duplicate queue of a, called b. Every write
	// into a will also be done on b. Each queue is allowed to
	// have at most ONE duplicate queue.
	dfbfix b(‘‘b’’);
	a.asDup(b);
	// Thus, when another duplicate is needed, you write is as
	dfbfix c(‘‘c’’);
	b.asDup(c);

[0169] During the communication of “dfix” objects, the queues keep track of some statistics on the values that are passed through it. You can use the “<<” operator and the member function “stattitle( )” to make these statistics visible.

[0170] The next program demonstrates these statistics

[0171]

	#include “qlib.h”
	void main( )
	{
	dfbfix a (“a”);
	a << dfix(2);
	a << dfix(l);
	a << dfix(3);
	a.stattitle(cout);
	cout << a;
	}

[0172] When running this program, the following appears on screen

[0173]

Name	put	get	MinVal	@idx	MaxVal	@idx	Max#	®idx
A	3	0	l.0000e+00	2	3.0000e+00	3	3	3

[0174] The first line is printed by the “stattitle( )” call as a mnemonic for the fields printed below. The next line is the result of passing the queue to the standard output stream object. The fields mean the following:

[0175]

	Name	The name of the queue
	put	The total number of elements “put ( )” into
		the queue
	get	The total number of elements “get ( )” from
		the queue
	MinVal	The lowest element put onto the queue
	@idx	The put sequential number that passed
		this lowest element
	MaxVal	The highest element put onto the queue
	@idx	The put sequential number that passed
		this highest element
	Max#	The maximal queue length that occurred
	@idx	The put sequential number that resulted ion
		this maximal queue length

[0176] Globals and Derivatives for dfbfix

[0177] There are two special derivates of “dfbfix”. Both are derived classes such that you can use them wherever you would use a “dfbfix”. Only the first will be discussed here, the other one is related to cycle-true simulation and is discussed in section “Faster Communications”.

[0178] The “dfbfix_nil” object is like a “/dev/null” drain. Every “dfix” written into this queue is thrown. A read operation from such a queue results in a runtime error.

[0179] There are two global variables related to queues. The “listOfFB” is a pointer to a list of queues, containing every queue object you have declared in your program. The member function call “nextFB( )” will return the successor of the queue in the global list. For example, the code snippet

[0180]

	dfbfix *r;
	for ( r = listOfFB ; r ; r = r->nextFB ( ) )
	{
	. . .
	}

will walk trough all the queues present in the OCAPI program.

[0181] The other global variable is “nilFB”, which is of the type “dfbfix_nil”. It is intended to be used as a global trashcan.

The Basic Block

[0182] OCAPI supports the dataflow simulation paradigm. In order to define the actors to the system, one “base” class is used, from which all actors will inherit. In order to do untimed simulations, one should follow a standard template to which new actor classes must conform. In this section, the standard template will be introduced, and the writing style is documented.

[0183] Basic Block Include and Code File

[0184] Each new actor in the system is defined with one header file and one source code C++ file. We define a standard block, “add”, which performs an addition.

[0185] The include file, “add.h”, looks like

[0186]

	#ifndef ADD_H
	#define ADD_H
	#include ‘‘qlib.h’’
	class add : public base
	{
	public:
	add (char *name, FB & _in1, FB & _in2, FB & _o1);
	int run( );
	private:
	FB *in1;
	FB *in2;
	FB *o1;
	};
	#endif

[0187] This defines a class “add”, that inherits from “base”. The “base” object is the one that OCAPI likes to work with, so you must inherit from it in order to obtain an OCAPI basic block.

[0188] The private members in the block are pointers to communication queues. Optionally, the private members should also contain state, for example the tap values in a filter. The management of state for untimed blocks is entirely the responsibility of the user; as far as OCAPI is concerned, it does not care what you use as extra variables.

[0189] The public members include a constructor and an execution call “run”. The constructor must at least contain a name, and a list of the queues that are used for communication. Optionally, some parameters can be passed, for instance in case of parametrized blocks (filters with a variable number of taps and the like).

[0190] The contents of the adder block will be described in “add.cxx”.

[0191]

	#include ‘‘add.cxx’’
	add::add(char *name, FB & _in1, FB & _in2, FB & _o1):
	base(name)
	{
	in1 = _in1.asSource(this);
	in2 = _in2.asSource(this);
	o1 = _o1.asSink (this);
	}
	int add::run( )
	{
	// firing rule
	if (in1->getSize( ) < 1)
	return 0;
	if (in2->getSize( ) < 1)
	return 0 ;
	o1->put (in1->get( ) + in2->get( ));
	return 1;
	}

[0192] The constructor passes the name of the object to the “base” class it inherits from. In addition, it initializes private members with the other parameters. In this example, the communication queue pointers are initialized. This is not done through simple pointer assignment, but through function calls “asSource” and “asSink”. This is not obligatory, but allows OCAPI to analyze the connectity in between the basic blocks. Since a queue is intended for point-to-point communication, it is an error to use a queue as input or ouput more then once. The function calls “asSource” and “asSink” keep track of which blocks source/sink which queues. They will return a runtime error in case a queue is sourced or sinked more then once. The constructor can optionally also be used to perform initialization of other private data (state for instance). The “run( )” method contains the operations to be performed when the block is invoked. The behavior is described in an iterative way. The “run” function must return an integer value, 1 if the block succeeded in performing the operation, and 0 if this has failed.

[0193] This behavior consists of two parts: a firing rule and an operative part. The firing rule must check for the availability of data on the input queues. When no sufficient data is present (checked with the “getSize( )” member call), it stops execution and returns 0. When sufficient data is present, execution can start. Execution of an untimed behavior can use the different C++ control constructs available. In this example, the contents of the two input queues is read, the result is added and put into the ouput queue. After execution, the value 1 is returned to signal the behavior has completed.

[0194] Predefined Standard Blocks: File Sources and Sinks

[0195] The OCAPI library contains three predefined standard blocks, which is a file source “src”, a file sink “snk”, and a ram storage block “ram”.

[0196] The file sources and sinks define operating system interfaces and allow you to bring file data into an OCAPI simulation, and to write out resulting data to a file. The examples below show various declarations of these blocks. Data in these files is formatted as floating point numbers separated by white space. For output, newlines are used as whitespace.

[0197]

	// define a file source block, with name a, that will read
	// data from the file ‘‘in.dat’’ and put it into the queue k
	dfbfix k(‘‘k’’);
	src a(‘‘a’’, k, ‘‘in.dat’’);
	// an alternative definition is
	dfbfix k(‘‘k’’);
	src a (‘‘a’’, k);
	a.setAttr(src::FILENAME, ‘‘in.dat’’);
	// which also gives you a complex version
	dfbfix k1(‘‘k1’’);
	dfbfix k2(‘‘k2’’);
	src a (‘‘a’’, k1, k2);
	a.setAttr (src::FILENAME, ‘‘in.dat’’);
	// define a sink block b, that will put data from queue o
	// into a file ‘‘out.dat’’.
	dfbfix o(‘‘o’’);
	snk b(‘‘b’’, o, ‘‘out.dat’’);
	// an alternative definition is
	dfbfix o(‘‘o’’);
	snk b(‘‘b’’, o);
	b.setAttr (snk::FILENAME, ‘‘out.dat’’);
	// which gives one also a complex version
	dfbfix o1(‘‘o1’’);
	dfbfix o2 (‘‘o2’’);
	snk b(‘‘b”, o1, o2);
	b.setAttr (snk::FILENAME, ‘‘out.dat’’);
	// the snk mode has also a matlab-goodie which will format
	// output data into a matrix A that can be read in directly
	// by Matlab.
	dfbfix o(‘‘o’’);
	snk b(‘‘b’’, o, ‘‘out.m’’);
	b.setAttr (snk::FILENAME, ‘‘out.m’’);
	b.setAttr(snk::MATLABMODE, 1);

[0198] Predefined Standard Blocks: RAM

[0199] The ram untimed block is intended to simulate single-port storage blocks at high level. By necessity, some interconnect assumptions had to be made on this block. On the other hand, it is supported all the way through code generation.

[0200] OCAPI does not generate RAM cells. However, it will generate appropriate connections in the resulting system netlist, onto which a RAM cell can be connected.

[0201] The declaration of a ram block is as follows.

[0202]

	// make a ram a, with an address bus, a data input bus, a
	// data output bus, a read command line, a write command
	// line, with 64 locations
	dfbfix address(‘‘address’’);
	dfbfix data_in(‘‘data_in’’);
	dfbfix data_out(‘‘data_out’’);
	dfbfix read_c(‘‘read_c’’);
	dfbfix write_c(‘‘write_c’’);
	ram a (‘‘a’’,address,data_in,data_out,write_c,read_c,64);
	// clear the ram
	a.clear( );
	// fill the ram with the linear sequence data = k1+address
	// * k2;
	a.fil(k1, k2);
	// dump the contents of a to cout
	a.show( );

[0203] The execution semantics of the ram are as follows. For each read or write, an address, a read command and a write command must be presented. If the read command equals “dfix(1)”, a read will be performed, and the value stored at the location presented through “address” will be put on “data_out”. If the read command equals any other value, a dummy byte will be presented at “data_out”. If no read command was presented, no data will be presented on “data_out”. For writes, an identical story holds for reads on the “data_in” input: whenever a write command is presented, the data input will be consumed. When the write command equals 1, then the data input will be stored in the location provided through “address”. When a read and write command are given at the same time, then the read will be performed before the write. The ram also includes an online “purifier” that will generate a warning message whenever data from an unwritten location is read.

Untimed Simulations

[0204] Given the descriptions of one or more untimed blocks, a simulation can be done. The description of a simulation requires the following to be included in a standard C++ “main( )” procedure:

• • [0205] The instantiation of one or more basic blocks.
  • [0206] The instantiation of one or more communication queues that interconnect the blocks
  • [0207] The setup of stimuli. Either these can be included at runtime by means of the standard file source blocks, or else dedicated C++ code can be written that fills up a queue with stimuli.
  • [0208] A schedule that drives the execution methods of the basic blocks.

[0209] A schedule, in general, is the specification of the sequence in which block firing rules must be tested (and fired if necessary) in order to run a simulation. There has been quite some research in determining how such a schedule can be constructed automatically from the interconnection network and knowledge of the block behavior. Up to now, an automatic mechanism for a general network with arbitrary blocks has not been found. Therefore, OCAPI relies on the designer to construct such a schedule.

[0210] Layout of an Untimed Simulation

[0211] In this section, the template of the standard simulation program will be given, along with a description of the “scheduler” class that will drive the simulation. A configuration with the “adder” block (described in the section on basic blocks) is used as an example.

[0212]

	#include ‘‘qlib.h”
	#include ‘‘add.h”
	void main( )
	{
	dfbfix i1(“i1”);
	dfbfix i2(“i2”);
	dfbfix o1(“o1”) ;
	src SRC1(“SRC1”, i1,“SRC1”);
	src SRC2(“SRC2”, i2, “SRC2”);
	add ADD (“ADD”, i1, 12, o1);
	snk SNK1(“SNK1”, o1, “SNK1”);
	schedule S1 (“S1”);
	S1.next(SRC1);
	S1.next(SRC2);
	S1.next(ADD );
	S1.next(SNK1);
	while (S1.run( )) ;
	i1.stattitle(cout);
	cout << i1;
	cout << i2;
	cout << o1;
	}

[0213] The simulation above instantiates three communication buffers, that interconnect four basic blocks. The instantiation defines at the same time the interconnection network of the simulation. Three of the untimed blocks are standard file sources and sinks, provided with OCAPI. The “add” block is a user defined one.

[0214] After the definition of the interconnection network, a schedule must be defined. A simulation schedule is constructed using “schedule” objects. In the example, one schedule object is defined, and the four blocks are assigned to it by means of a “next( )” member call.

[0215] The order in which “next( )” calls are done determines the order in which firing rules will be tested. For each execution of the schedule object “S1” the “run( )” methods of “SRC1”, “SRC2”, “ADD” and “SNK1” are called, in that order. The execution method of a scheduler object is called “run( )”. This function returns an integer, equal to one when at least on block in the current iteration has executed (i.e. the “run( )” of the block has returned one). When no block has executed, it returns zero.

[0216] The while loop in the program therefore is an execution of the simulation. Let us assume that the directory of the simulator executable contains the two required stimuli files, “SRC1” and “SRC2”. Their contents is as follows

[0217]

	SRC1	SRC2	-- not present in the file
	----	----	-- not present in the file
	1	4
	2	5
	3	6

[0218] When compiling and running this program, the simulator responds:

• [0219] *** INFO: Defining block SRC1
• [0220] *** INFO: Defining block SRC2
• [0221] *** INFO: Defining block ADD
• [0222] *** INFO: Defining block SNK1

[0223]

Name	put	get	MinVal	@idx	MaxVal	@idx	Max#	@idx
i1	3	3	l.0000e+00	1	3.0000e+00	3	1	1
i2	3	3	4.0000e+00	1	6.0000e+00	3	1	1
o1	3	3	5.0000e+00	1	9.0000e+00	3	1	1

and in addition has created a file “SNK1”, containing

• [0224] SNK1—not present in the file
• [0225] ----—not present in the file
• [0226] 5.000000e+00
• [0227] 7.000000e+00
• [0228] 9.000000e+00

[0229] The “INFO” message appearing on standard output are a side effect of creating a basic block. The table at the end is produced by the print statements at the end of the program.

[0230] More on Schedules

[0231] If you would examine closely which blocks are fired in which iteration, (for instance with a debugger) then you would find

[0232]

	iteration 1
	run SRC1 => i1 contains 1.0
	run SRC2 => i2 contains 4.0
	run ADD => o1 contains 5.0
	run SNK1 => write out o1
	schedule.run( ) returns 1
	iteration 2
	run SRC1 => i1 contains 2.0
	run SRC2 => i2 contains 5.0
	run ADD => o1 contains 7.0
	run SNK1 => write out o1
	schedule.run( ) returns 1
	iteration 3
	run SRC1 => i1 contains 3.0
	run SRC2 => i2 contains 6.0
	run ADD => o1 contains 9.0
	run SNK1 => write out o1
	schedule.run( ) returns 1
	iteration 4
	run SRC1 => at end-of-file, fails
	run SRC2 => at end-of-file, fails
	run ADD => no input tokens, fails
	run SNK1 => no input tokens, fails
	schedule.run( ) returns 0 => end simulation

[0233] There are two schedule member functions, “traceOn( )” and “traceOff( )”, that will produce similar information for you. If you insert

• [0234] S.traceOn( );
  just before the while loop, then you see
• [0235] *** INFO: Defining block SRC1
• [0236] *** INFO: Defining block SRC2
• [0237] *** INFO: Defining block ADD
• [0238] *** INFO: Defining block SNK1
• [0239] S1 [SRC1 SRC2 ADD SNK1]
• [0240] S1 [SRC1 SRC2 ADD SNK1]
• [0241] S1 [SRC1 SRC2 ADD SNK1]
• [0242] S1 [ ]

[0243]

Name	put	get	MinVal	@idx	MaxVal	@idx	Max#	@idx
i1	3	3	l.0000e+00	1	3.0000e+00	3	1	1
i2	3	3	4.0000e+00	1	6.0000e+00	3	1	1
o1	3	3	5.0000e+00	1	9.0000e+00	3	1	1

appearing on the screen. This trace feature is convenient during schedule debugging.

[0244] In the simulation ouput, you can also notice that the maximum number of tokens in the queues never exceeds one. When you had entered another schedule sequence, for example

[0245]

	schedule S1(“S1”);
	S1.next(ADD );
	S1.next(SRC2);
	S1.next(SRC1);
	S1.next(SNK1);

then you would notice that the maximum number of tokens on the queues would result in different figures. On the other hand, the resulting data file, “SNK1”, will contain exactly the same results. This demonstrates one important property of dataflow simulations: any arbitrary but consistent schedule yields the same results. only the required amount of storage will change from schedule to schedule.

[0246] In multirate systems, it is convenient to have different schedule objects and group all blocks working on the same rate in one schedule.

[0247] Profiling in Untimed Simulations

[0248] Untimed simulations are not targeted to circuit implementation. Rather, they have an explorative character. Besides the queue statistics, OCAPI also enables you to do precise profiling of operations. The requirement for this feature is that

• • [0249] You use “schedule” objects to construct the simulation
  • [0250] You describe block behavior with “dfix” objects

[0251] Profiling is by default enabled. To view profiling results, you send the schedule object under consideration to the standard output stream. In the “main” example program given above, you can modify this as

[0252]

	include ‘‘qlib.h’’
	include ‘‘add.h’’
	void main( )
	{
	. . .
	schedule S1(“S1”);
	. . .
	cout << S1;
	}

[0253] When running the simulation, you will see the following appearing on stdout:

• [0254] *** INFO: Defining block SRC1
• [0255] *** INFO: Defining block SRC2
• [0256] *** INFO: Defining block ADD
• [0257] *** INFO: Defining block SNK1

[0258]

Name	put	get	MinVal	@idx	MaxVal	@idx	Max#	@idx
i1	3	3	1.0000e+00	1	3.0000e+00	3	1	1
i2	3	3	4.0000e+00	1	6.0000e+00	3	1	1
o1	3	3	5.0000e+00	1	9.0000e+00	3	1	1

[0259] Schedule S1 ran 4 times:

[0260]

	SRC1	3
	SRC2	3
	ADD	3
	+	3
	SNK1	3

[0261] For each schedule, it is reported how many times it was run. Inside each schedule, a firing count of each block is given. Inside each block, an operation execution count is given. The simple “add” block gives the rather trivial result that there were three additions done during the simulation.

[0262] The gain in using operation profiling is to estimate the computational requirement for each block. For instance, if you find that you need to do 23 multiplications in a block that was fired 5 times, then you would need at least five multipliers to guarantee the block implementation will need only one cycle to execute.

[0263] Finally, if you want to suppress operation profiling for some blocks, then you can use the member function call “noOpsCnt( )” for each block. For instance, writing

• [0264] ADD.noOpsCnt( );
  suppresses operation profiling in the ADD block.
  Implementation

[0265] The features presented in the previous sections contain everything you need to do untimed, high level simulations. These kind of simulations are useful for initial development. For real implementation, more detail has to be added to the descriptions.

[0266] OCAPI makes few assumptions on the target architecture of your system. One is that you target bitparallel and synchronous hardware. Synchronicity is not a basic requirement for OCAPI. The current version however constructs single-thread simulations, and also assumes that all hardware runs at the same clock. If different clocks need to be implemented, then a change to the clock-cycle true simulation algorithm will have to be made. Also, it is assumed that one basic block will eventually be implemented into one processor.

[0267] One question that comes to mind is how hardware sharing between different basic blocks can be expressed. The answer is that you will have to construct a basic block that merges the two behaviors of two other blocks. Some designers might feel reluctant to do this. On the other hand, if you have to write down merged behavior, you will also have to think about the control problems that are induced from doing this merging. OCAPI will not solve this problem for you, though it will provide you with the means to express it.

[0268] Before code generation will translate a description to an HDL, one will have to take care of the following tasks:

• • [0269] One will have to specify wordlengths. The target hardware is capable of doing bitparallel, fixed point operations, but not of doing floating point operations. One of the design tasks is to perform the quantisation on floating point numbers. The “dfix” class discussed earlier contains the mechanisms for expressing fixed point behavior.
  • [0270] One will have to construct a clock-cycle true description. In constructing this description, one will not have to allocate actual hardware, but rather express which operations one expects to be performed in which clock cycle. The semantical model for describing this is clock cycle true behavior consists of a finite state machine, and a set of signal flow graphs. Each signal flow graph expresses one cycle of implemented behavior. This style of description splits the control operations from data operations in your program. In contrast, the untimed description you have used before has a common representation of control and data.

[0271] OCAPI does not force an ordening on these tasks. For instance, one might first develop a clock cycle true description on floating point numbers, and afterwards tackle the quantization issues. This eases verification of the clock-cycle true circuit to the untimed high level simulation.

[0272] The final implementation also assumes that all communication queues will be implemented as wiring. They will contain no storage, nor they will be subject to buffer synthesis. In a dataflow simulation, initial buffering values can however be necessary (for instance in the presence of feedback loops). In OCAPI, such a buffer must be implemented as an additional processor that incorporates the required storage. The resulting system dataflow will become deadlocked because of this. The cycle scheduler however, that simulates timed descriptions, is clever enough to look for these ‘initial tokens’ inside of the descriptions.

[0273] In the next sections, the classes that allow you to express clock cycle true behavior are introduced.

Signals and Signal Flowgraphs

[0274] Some initial considerations on signals are introduced first.

[0275] Hardware Versus Software

[0276] Software programs always use memory to store variables. In contrast, hardware programs work with signals, which might or might not be stored into a register. This feature can be expressed in OCAPI by using the “_sig” class. Simply speaking, a “_sig” is a “dfix” for which one has indicated whether is needs storage or not.

[0277] In implementation, a signal with storage is mapped to a net driven by a register, while an immediate signal is mapped to a net driven by an operator.

[0278] Besides the storage issue, a signal also departs from the concept of “scope” one uses in a program. For instance, in a function one can use local variables, which are destroyed (i.e. for which the storage is reclaimed) after one has executed the function. In hardware however, one controls the signal-to-net mapping by means of the clock signal.

[0279] Therefore one have to manage the scope of signals. The signal scope is expressed by using a signal flowgraph object, “sfg”. A signal flowgraph marks a boundary on hardware behavior, and will allow subsequent synthesis tools to find out operator allocation, hardware sharing and signal-to-net mapping.

[0280] The _sig Class and Related Operations

[0281] Hardware signals can expressed in three flavors. They can be plain signals, constant signals, or registered signals. The following example shows how these three can be defined.

[0282]

	// define a plain signal a, with a floating point dfix
	// inside of it.
	_sig a(‘‘a’’);
	// define a plain signal b, with a fixed point dfix inside
	// of it.
	_sig b(‘‘b’’, dfix(0,10,8));
	// define a registered signal c, with an initial value k
	// and attached to a clock ck.
	dfix k(0.5);
	clk ck;
	_sig c(‘‘c’’, ck, k);
	// define a constant signal d, equal to the value k
	_sig d(k);

[0283] The registered signals, and more in particular the clock object, are explained more into detail when signal flowgraphs and finite state machines are discussed. This section concentrates on operations that are available for signals.

[0284] Using signals and signal operations, one can construct expressions. The signal operations are a subset of the operations on “dfix”. This is because there is a hardware operator implementation behind each of these operations.

• • [0285] +, −, *
    • [0286] Standard addition, subtraction (including unary minus), multiplication
  • [0287] &, |, ^, ˜
    • [0288] Bitwise and, or, exor, and not operators
  • [0289] ==, !=, <=, >=, <, >
    • [0290] Relational operators
  • [0291] <<, >>
    • [0292] Left and right shifts
  • [0293] s.cassign(s1,s2)
    • [0294] Conditional assignment with s1 or s2 depending on s
  • [0295] cast(T,s)
    • [0296] Convert the type of s to the type expressed in “dfix” T
  • [0297] lu(L,s)
    • [0298] Use s as in index into lookuptable L and retrieve
  • [0299] msbpos(s)
    • [0300] Return the position of the msb in s

[0301] Precision considerations are the same as for “dfix”. That is, precision is at most the mantissa precision of a double (53 bits). For the bitwise operations, 32 bits are assumed (a long). “cast”, “lu” and “msbpos” are not member but friend functions. In addition, “msbpos” expects fixed-point signals.

[0302]

	_sig a(‘‘a’’);
	_sig b(‘‘b’’);
	_sig c(‘‘c’’);
	// some simple operations
	c = a + b;
	c = a − b;
	c = a * b;
	// bitwise operations works only on fixed point signals
	_sig e(dfix(0xff, 10, 0));
	_sig d(‘‘d’’,dfix(0,10,0));
	_sig f(‘‘f’’ ,dfix(0,10,0));
	f = d & e;
	f = d | e;
	f = ˜d;
	f = d {circumflex over ( )} _sig(dfix(3,10,0));
	// shifting
	// a dfix is automatically promoted to a constant _sig
	f = d << dfix (3,8,0);
	// conditional assignment
	f = (d < dfix(2,10,0)).cassign(e,d);
	// type conversion is done with cast
	_sig g(‘‘g’’,dfix(0,3,0));
	g = cast (dfix(0,3,0), d);
	// a lookup table is an array of unsigned long
	unsigned long j = {1, 2, 3, 4, 5};
	// a lookuptable with 5 elements, 3 bits wide
	lookupTable j_lookup (‘‘j_lookup’’, 5, dfix (0,3,0)) = j;
	// find element 2
	g = lu(j_lookup, dfix (2,3,0));

[0303] If one is interested in simulation only, then one should not worry too much about type casting and the like. However, if one intends implementation, then some rules are at hand. These rules are induced by the hardware synthesis tools. If one fails to obey them, then one will get a runtime error during hardware synthesis.

• • [0304] All operators, apart from multiplication, return a signal with the same wordlength as the input signal.
  • [0305] Multiplication returns a wordlength that is the sum of the input wordlengths.
  • [0306] Addition, subtraction, bitwise operations, comparisons and conditional assignment require the two input operands to have the same wordlength.

[0307] Some common pitfalls that result of this restriction are the following.

• • [0308] Intermediate results will, by default, not expand wordlength. In contrast, operations on dfix do not loose precision on intermediate results. For example, shifting an 8 bit signal up 8 positions will return you the value of zero, on 8 bits. If you want too keep up the precision, then you must first cast the operation to the desired output wordlength, before doing the shift.
  • [0309] The multiplication operator increases the wordlength, which is not automatically reduced when you assign the result to a signal of smaller with. If you want to reduce wordlength, then you must do this by using a cast operation.

[0310] For complex expressions, these type promotion rules look a bit tedious. They are however used because they allow you to express behavior precisely downto the bit level. For example, the following piece of code extracts each of the bits of a three bit signal:

• [0311] _sig threebits(dfix(6,3,0));
• [0312] dfix bit(0,1,0);
• [0313] _sig bit2(“bit2”), bit1(“bit1”), bit0(“bit0”);
• [0314] bit2=cast(bit, threebits>>dfix(2));
• [0315] bit1=cast(bit, threebits>>dfix(1));
• [0316] bit0=cast(bit, threebits);

[0317] These bit manipulations were not possible without the given type promotion rules.

[0318] For hardware implementation, the following operators are present.

• • [0319] Addition and subtraction are implemented on ripple-carry adder/subtractors.
  • [0320] Multiplication is implemented with a booth multiplier block.
  • [0321] Casts are hardwired.
  • [0322] Shifts are either hardwired in case of constant shifts, or else a barrel shifter is used in case of variable shifts.
  • [0323] Comparisons are implemented with dedicated comparators (in case of constant comparisons), or subtractions (in case of variable comparisons).
  • [0324] Bitwise operators are implemented by their direct gate equivalent at the bit level.
  • [0325] Lookup tables are implemented as PLA blocks that are mapped using two-level or multi-level random logic.
  • [0326] Conditional assignment is done using multiplexers.
  • [0327] Msbit detection is done using a dedicated msbit-detector.

[0328] Globals and Utility Functions for Signals

[0329] There are a number of global variables that directly relate to the “_sig” class, as well as the embedded “sig” class. In normal circumstances, you do not need to use these functions.

[0330] The variables “glbNumberOf_Sig” and “glbNumberOfSig” contain the number of “_sig” and “sig” that your program has defined. The variable “glbNumberOfReg” contains the number of “sig” that are of the register type. This represents the word-level register count of your design. The “glbSigHashConflicts” contain the number of hash conflicts that are present in the internal signal data structure organization. If this number is more then, say 5% of “glbNumberOf_Sig”, then you might consider knocking at OCAPIs complaint counter. The simulation is not bad if you exceed this bound, only it will go slower.

[0331] The variable “glbListOfSig” contains a global list of signals in your system. You can go through it by means of

[0332]

	sig *run;
	for (run = glbListOfSig; run; run = run−>nextsig( ))
	{
	. . .
	}

[0333] For each such a “sig”, you can access a number of utility member functions.

• • [0334] “isregister( )” returns 1 when a signal is a register.
  • [0335] “isconstant( )” returns 1 when a signal is a constant value.
  • [0336] “isterm( )” returns 1 when you have defined this signal yourself. These are signals which are introduced through “_sig( )” class constructors. OCAPI however also adds signals of its own.
  • [0337] “getname( )” returns the “char *” name you have used to define the signal.
  • [0338] “get_showname( )” returns the “char *” name of the signal that is used for code generation. This is equal to the original name, but with a unique suffix appended to it.

[0339] The sfg Class

[0340] In order to construct a timed (clocked) simulation, signals and signals expressions must be assigned to a signal flowgraph. A signal flowgraph (in the context of OCAPI) is a container that collects all behavior that must be executed during one clock cycle.

[0341] The sfg behavior contains

• • [0342] A set of expressions using signals
  • [0343] A set of inputs and outputs that relate signals to output and input queues

[0344] Thus, a signal flowgraph object connects local behavior (the signals) to the system through communications queues. In hardware, the indication of input and output signals also results in ports on your resulting circuit.

[0345] A signal flowgraph can be a marker of hardware scope. This is also demonstrated by the following example.

[0346]

	_sig a(‘‘a’’);
	_sig b(‘‘b’’);
	_sig c(dfix(2));
	dfbfix A(‘‘A’’);
	dfbfix B(‘‘B’’);
	// a signal flowgraph object is created
	sfg add_two, add_three;
	// from now on, every signal expression written down will
	// be included in the signal flowgraph add_two
	add_two.starts( );
	a = b + c;
	// You must also give a name to add_two, for code
	// generation
	add_two << ‘‘add_two’’;
	// also, inputs and ouputs have to be indicated.
	// you use the input and ouput objects ip and op for this
	add_two << ip(b, B);
	add_two << op(a, A);
	// next expression will be part of add three
	add_three.starts( );
	a = b + dfix(3);
	add_three << ‘‘add_three’’;
	add_three << ip(b,B);
	add_three << op(a,A);
	// you can also to semantical checks on signal flowgraphs
	add_two.check( );
	add_three.check( );

[0347] The semantical check warns you for the following specification errors:

• • [0348] Your signal flowgraph contains a signal which is not declared as a signal flowgraph input and at the same time, it is not a constant or a register. In other words, your signal flowgraph has a dangling input.
  • [0349] You have written down a combinatorial loop in your signal flowgraph. Each signal must be ultimately dependent on registered signals, constants, or signal flowgraph inputs. If any other dependency exists, you have written down a combinatorial loop for which hardware synthesis is not possible.
    Execution of a Signal Flowgraph

[0350] A signal flowgraph defines one clock cycle of behavior. The semantics of a signal flowgraph execution are well defined.

• • [0351] At the start of an execution, all input signals are defined with data fetched from input queues.
  • [0352] The signal flowgraph output signals are evaluated in a demand driven way. That is, if they are defined by an expression that has signal operands with known values, then the ouput signal is evaluated. Otherwise, the unknown values of the operands are determined first. It is easily seen that this is a recursive process. Signals with known values are: registered signals, constant signals, and signals that have already been calculated in the current execution.
  • [0353] The execution ends by writing the calculated output values to the output queues.

[0354] Signal flowgraph semantics are somewhat related to untimed blocks with firing rules. A signal flowgraph needs one token to be present on each input queue. Only, the firing rule on a signal flowgraph is not implemented. If the token is missing, then the simulation crashes. This is a crude way of warning you that you are about to let your hardware evaluate a nonsense result.

[0355] The relation with untimed block firing rules will allow to do a timed simulation which consist partly of signal flowgraph descriptions and partly of untimed basic blocks. The section “Timed simulations will treat this more into detail.

[0356] Running a Signal Flowgraph by Hand

[0357] A signal flowgraph is only part of a timed description. The control component (an FSM) still needs to be introduced. There can however be situations in which you would like to run a signal flowgraph directly. For instance, in case you have no control component, or if you have not yet developed a control description for it.

[0358] The “sfg” member function “run( )” performs the execution of the signal flowgraph as described above. An example is used to demonstrate this.

[0359]

	#include “qlib.h”
	void main( )
	{
	_sig a(“a”);
	_sig b(“b”);
	_sig c(dfix(2));
	dfbfix A(“A”);
	dfbfix B(“B”);
	sfg add_two;
	add two.starts( );
	a = b + c;
	add_two << “add_two”;
	add_two << ip(b, B);
	add_two << op(a, A);
	add_two.check( );
	B << dfix(1) << dfix(2);
	// running silently
	add_two.eval( );
	cout << A.get( ) << “\n”;
	// running with debug information
	add_two.eval(cout);
	cout << A.get( ) << “\n”;
	add_two.eval(cout);
	}

[0360] When running this simulation, the following appears on the screen.

[0361]

	3.000000e+00
	add_two(	b	2)
	:	a	4
	=>	a	4
	4.000000e+00
	add_two(Queue Underflow @ get in queue B

[0362] The first line shows the result in the first “eval( )” call. When this call is given an output stream as argument, some additional information is printed during evaluation. For each signal flowgraph, a list of input values is printed. Intermediate signal values are printed after the “:” at the beginning of the line. The output values as they are entered in the ouput queues are printed after the “=>”. Finally, the last line shows what happens when “eval( )” is called when no inputs are available on the input queue “B”.

[0363] For signal flowgraphs with registered signals, you must also control the clock of these signals. An example of an accumulator is given next.

[0364]

	#include “qlib.h”
	void main( )
	{
	clk ck;
	_sig a(“a”,ck,dfix(0));
	_sig b(“b”);
	dfbfix A(“A”);
	dfbfix B(“B”);
	sfg accu;
	accu.starts( );
	a = a + b;
	accu << “accu”;
	accu << ip(b, B);
	accu << op(a, A);
	accu.check( );
	B << dfix(1) << dfix(2) << dfix(3);
	while(B.getSize( ))
	{
	accu.eval(cout);
	accu.tick(ck);
	}
	}

[0365] The simulation is controlled in a while loop that will consume all input values in queue “B”. After each run, the clock attached to registered signal “a” is triggered. This is done indirectly through the “sfg” member call “tick( )”, that updates all registered signals that have been assigned within the scope of this “sfg”. Running this simulation results in the following screen ouput

[0366]

	accu	(	b	1)
		:	a	0/	1
		=>	a	0/	1
	accu	(	b	2)
		:	a	1/	3
		=>	a	1/	3
	accu	(	b	3)
		:	a	3/	6
		=>	a	3/	6

[0367] The registered signal “a” has two values: a present value (shown left of “/”), and a next value (shown right of “/”). When the clock ticks, the next value is copied to the present value. At the end of the simulation, registered signal “a” will contain 6 as its present value. The ouput queue “A” however will contain the 3, the “present value” of “a” during the last iteration.

[0368] Finally, if you want to include a signal flowgraph in an untimed simulation, you must make shure that you implement a firing rule that guards the sfg evaluation.

[0369] An example that incorporates the accumulator into an untimed basic block is the following.

[0370]

	#include “qlib.h”
	class accu : public base
	{
	public:
	accu(char *name, dfbfix &i, dfbfix &o);
	int run( );
	private:
	dfbfix *ipq;
	dfbfix *opq;
	sfg _accu;
	clk ck;
	}
	accu::accu(char *name, dfbfix &i, dfbfix &o) : base(name)
	{
	ipq = i.asSource(this);
	opq = o.asSink(this);
	_sig a(“a”,ck,dfix(0));
	_sig b(“b”);
	_accu.starts( );
	a = a + b;
	_accu << “accu”;
	_accu << ip(b, *ipq);
	_accu << op(a, *opq);
	_accu.check( );
	}
	int accu::run( )
	{
	if (ipq->getSize( ) < 1)
	return 0;
	_accu.eval( );
	accu.tick(ck);
	}

[0371] In this example, the signal flowgraph _accu is included into the private members of class _accu.

[0372] Globals and Utility Functions for Signal Flowgraphs

[0373] The global variable “glbNumberOfSfg” contains the number of “sfg” objects that you have constructed in your present OCAPI program. Given an “sfg( )” object, you have also a number of utility member function calls.

• • [0374] “getname( )” returns the “char *” name of the signal flowgraph.
  • [0375] “merge( )” joins two signal flowgraphs.
  • [0376] “getisig(int n)” returns a “sig *” that indicates which signal corresponds to input number “i” of the signal flowgraph. If 0 is returned, this input does not exist.
  • [0377] “getiqueue(int n)” returns the queue (“dfbfix *”) assigned to input number “i” of the signal flowgraph. If 0 is returned, then this input does not exist.
  • [0378] “getosig(int n)” returns a “sig *” that indicates which signal corresponds to output number “i” of the signal flowgraph. If 0 is returned, this output does not exist.
  • [0379] “getoqueue(int n)” returns the queue (“dfbfix *”) assigned to output number “i” of the signal flowgraph. If 0 is returned, then this output does not exist.

[0380] You should keep in mind that a signal flowgraph is a data structure. The source code that you have written helps to build this data structure. However, a signal flowgraph is not executed by running your source code. Rather, it is interpreted by OCAPI. You can print this data structure by means of the “cg(ostream)” member call.

• [0381] For example, if you appended
• [0382] accu.cg(cout);

[0383] to the “running-an-sfg-by-hand” example, then the following output would be produced:

[0384]

	sfg accu
	inputs	{ b_2 }
	outputs	{ a_1 }
	code {
	a_1 = a_1_at1 + b_2;
	};

Finite State Machines

[0385] With the aid of signals and signal flowgraphs, you are able to construct clock-cycle true data processing behavior. On top of this data processing, a control sequencing component can be added. Such a controller allows to execute signal flowgraphs conditionally. The controller is also the anchoring point for true timed system simulation, and for hardware code generation. A signal flowgraph embedded in an untimed block cannot be translated to a hardware processor: you have to describe the control component explicitly.

The ctlfsm and State Classes

[0386] The controller model currently embedded in OCAPI is a Mealy-type finite state machine. This type of FSM selects the transition to the next state based on the internal state and the previous output value.

[0387] In an OCAPI description, you use a “ctlfsm” object to create such a controller. In addition, you make use of “state” objects to model controller states. The following example shows the use of these objects.

[0388]

	#include ‘‘qlib.h’’
	void main( )
	{
	sfg dummy;
	dummy << ‘‘dummy’’;
	// create a finite state machine
	ctlfsm f;
	// give it a name
	f << ‘‘theFSM’’;
	// create 2 states for it
	state rst;
	state active;
	// give them a name
	rst << ‘‘rst’’;
	active << ‘‘active’’;
	// identify rst as the initial state of
	// ctlfsm f
	f << deflt(rst);
	// identify active as a plain state of ctlfsm
	// f
	f << active;
	// create an unconditional transition from
	// rst to active
	rst << allways << active;
	// allways' is a historical typo and will be
	// replaced by “always” in the future
	// create an unconditional transition from
	// active to active, executing the dummy sfg.
	active << allways << dummy << active;
	// show what's inside f
	cout << f;
	}

[0389] There are two states in this fsm, “rst” and “active”. Both are inserted in the fsm by means of the “<<” operator. In addition, the “rst” state is identified as the default state of the fsm, by embedding it into the “deflt” object. An fsm is allowed to have one default state. When the fsm is simulated, then the state at the start of the first clock cycle will be “rst”. In the hardware implementation, a “reset” pin will be added to the processor that is used to initialize the fsm's state register with this state.

[0390] Two transitions are defined. A transition is written according to the template: starting state, conditions, actions, target state, all of this separated by the “<<” operator. The condition “allways” is a default condition that evaluates to true. It is used to model unconditional transitions.

[0391] The last line of the example shows a simple operation you can do with an fsm. By relating it to the output stream, the following will appear on the screen when you compile and execute the example.

[0392]

	digraph g
	{
	rst [shape=box];
	rst−>active;
	active−>active;
	}

[0393] This output represent a textual format of the state transition diagram. The format is that of the “dotty” tool, which produces a graphical layout of your state transition diagram.

[0394] “dotty” is commercial software available from AT&T.

[0395] You cannot simulate a “ctlfsm” object on itself. You must do this indirectly through the “sysgen” object, which is introduced in the section “Timed Simulations”.

[0396] The cnd Class

[0397] Besides the default condition “allways”, you can use also boolean expressions of registered signals. The signals need to be registered because we are describing a Mealy-type fsm. You construct conditions through the “cnd” object, as shown in the next example.

[0398]

	#include “qlib.h”
	void main( )
	{
	clk ck;
	_sig a(“a”,ck, dfix(0));
	_sig b(“b”,ck, dfix(0));
	_sig a_input(“a”);
	_sig b_input(“a”);
	dfbfix A(“A”);
	dfbfix B(“B”);
	sfg some_operation;
	// some operations go here . . .
	sfg readcond;
	readcond.starts( );
	a = a_input;
	b = b_input;
	readcond << “readcond”;
	readcond << ip(a_input,A);
	readcond << ip(binput,B);
	readcond.check( );
	// create a finite state machine
	ctlfsm f;
	f << “theFSM”;
	state rst;
	state active;
	state wait;
	rst	<< “rst”;
	active	<< “active”;
	wait	<< “wait”;
	f << deflt(rst);
	f << active;
	f << wait;
	rst	<< allways << readcond << active;
	active	<< _cnd(a) << readcond << some_operation
	<< wait;
	wait	<< (_cnd(a) && _cnd(b)) << readcond
	<< wait;
	wait	<<(!_cnd(a) ||!_cnd (b)) <<readcond<< active;
	}

[0399] A FAQ is why condition signals must be registers, and whether they can be plain signals also. The answer is simple: no, they can't. The fsm control object is a stand-alone machine that must be able to ‘boot’ every clock cycle. During one execution cycle, it will first select the transition to take (based on conditions), and then execute the signal flowgraphs that are attached to this transition. If “immediate” transition conditions had to be expressed, then the signals should be read in before the fsm transition is made, which is not possible: the execution of an sfg can only be done when a transition is selected, in other words: when the condition signals are known. Besides this semantical consideration, the registered-condition requirement will also prevent you from writing combinatorial control loops at the system level.

[0400] The first signal flowgraph “readcond” takes care of reading in two values “a” and “b” that are used in transition conditions. The sfg reads the signals “a” and “b” in through the intermediate signals “a_input” and “b_input”. This way, “a” and “b” are explicitly assigned in the signal flowgraph, and the semantical check “readcond.check( )” will not complain about unassigned signals.

[0401] The fsm below it defines three states. Besides an initial state “rst” and an operative state “active”, a wait state “wait” is defined, that is entered when the input signal “a” is high. This is expressed by the “_cnd(a)” transition condition in the second fsm transition. You must use “_cnd( )” instead of “cnd( )” because of the same reason that you must use “_sig( )” instead of “sig( )”: The underscore-type classes are empty boxes that allocate the objects that do the real work for you. This allocation is dynamic and independent of the C++ scope.

[0402] Once the wait state is entered, it can leave it only when the signals “a” or “b” go low. This is indicated in the transition condition of the third fsm transition. A “&&” operator is used to express the and condition. If the signals “a” and “b” remain high, then the wait state is not left. The transition condition of the last transition expresses this. It uses the logical not “!” and logical or “||” operators to express this.

[0403] The “readcond” signal flowgraph is executed at all transitions. This ensures that the signals “a” and “b” are updated every cycle. If you fail to do this, then the value of “a” and “b” will not change, potentially creating a deadlock.

[0404] To summarize, you can use either “always” or a logical expression of “_cnd( )” objects to express a transition condition. The signals use in the condition must be registers. This results in a Mealy-type fsm description

[0405] Utility Functions for fsm Objects

[0406] A number of utility functions on the “ctlfsm” and “state” classes are available for query purposes. This is only minimal: The objects are intended to be manipulated by the cycle scheduler and code generators.

[0407]

	sfg action;
	ctlfsm f;
	state s1;
	state s2;
	f << deflt(s1);
	f << s2;
	s1 << allways << s2;
	s2 << allways << action << s1;
	// run through all the state in f
	statelist *r;
	for (r = f.first; r; r = r−>next)
	{
	. . .
	}
	// print the nuymber of states in f,
	// print the number of transitions in f,
	// print the name of f,
	// print the number of sfg's in f
	cout << f.numstates( ) << ‘‘\n’’;
	cout << f.numtransitions( ) << ‘‘\n’’;
	cout << f.getname( ) << ‘‘\n’’;
	cout << f.numactions( ) << ‘‘\n’’;
	// print the name of a state
	cout << s1.getname( ) << ‘‘\n’’;

[0408] The Basic Block for Timed Simulations

[0409] Using signals, signal flowgraphs, finite state machines and states, you can construct a timed description of a block. Having obtained such a description, it is convenient to merge it with the untimed description. This way, you will have one class that allows both timed and untimed simulation. Of course, this merging is a matter of writing style, and nothing forces you to actually have both a timed and untimed description for a block.

[0410] The basic block example, that was introduced in the section “The basic block”, will now be extended with a timed version. As before, both an include file and a code file will be defined. The include file, “add.h”, looks like the following code.

[0411]

	#ifndef ADD_H
	#define ADD_H
	#include ‘‘qlib.h’’
	class add : public base
	{
	public:
	add(char *name, FB & _in1, FB & _in2, FB & _o1) ;
	// untimed
	int run( );
	// timed
	void	define( );
	ctlfsm  &fsm( ) {return _fsm};
	private:
	FB *in1;
	FB *in2;
	FB *o1;
	ctlfsm  _fsm;
	sfg _add;
	state  _go;
	};
	#endif

[0412] The private members now also contain a control fsm object, in addition to signal flowgraph objects and states. If you feel this is becoming too verbose, you will find help in the section “Faster description using macros”, that defines a macro set that significantly accelerates description entry.

[0413] In the public members, two additional member functions are declared: the “define( )” function, which will setup the timed description data structure, and the “fsm( )”, which returns a pointer to the fsm controller. Through this pointer, OCAPI accesses everything it needs to do simulations and code generation.

[0414] The contents of the adder block will be described in “add.cxx”.

[0415]

	#include ‘‘add.h’’
	add::add(char *name, FB & _in1, FB & _in2, FB & _o1) :
	base(name)
	{
	in1 = _in1.asSource(this);
	in2 = _in2.asSource(this);
	o1 = _o1.asSink (this);
	define( );
	}
	int add::run( )
	{
	. . .
	}
	void add::define( )
	{
	_sig i1(‘‘i1’’);
	_sig i2(‘‘i2’’);
	_sig ot(‘‘ot’’);
	_add << ‘‘add’’;
	_add.starts( );
	ot = i1 + i2;
	_add << ip(i1, *in1);
	_add << ip(i2, *in2);
	_add << op(ot, *o1);
	_fsm << ‘‘fsm’’;
	_go << ‘‘go’’;
	_fsm << deflt(_go);
	_go << allways << _add << _go;
	}

[0416] If the timed description, uses also registers, then a pointer to the global clock must also be provided (OCAPI generates single-clock, synchronous hardware). The easiest way is to extend the constructor of “add” with an additional parameter “clk &ck”, that will also be passed to the “define” function.

Timed Simulations

[0417] By obtaining timed descriptions for you untimed basic block, you are now ready to proceed to a timed simulation. A timed simulation differs from an untimed one in that it proceeds clock cycle by clock cycle. Concurrent behavior between different basic blocks is simulated on a cycle-by-cycle basis. In contrast, in an untimed simulation, this concurrency is present on an iteration by iteration basis.

[0418] The sysgen Class

[0419] The “sysgen” object is for timed simulations the equivalent of a “scheduler” object for untimed simulations. In addition, it also takes care of code and testbench generation, which explains the name.

[0420] The sysgen class is used at the system level. The timed “add” class, defined in the previous section, is used as an example to construct a system which uses untimed file sources and sinks, and a timed “add” class.

[0421]

	#include ‘‘qlib.h’’
	#include ‘‘add.h’’
	void main( )
	{
	dfbfix i1(“i1”);
	dfbfix i2(“i2”);
	dfbfix o1(“o1”);
	src SRC1(“SRC1”, i1, “SRC1”);
	src SRC2(“SRC2”, i2,“SRC2”);
	add ADD (“ADD”, i1, i2, o1);
	snk SNK1(“SNK1”, o1, “SNK1”);
	sysgen S1(“S1”);
	S1 << SRC1;
	S1 << SRC2;
	S1 << ADD.fsm( );
	S1 << SNK1;
	S1.setinfo(verbose);
	clk ck;
	int i;
	for (i=0; i<3; i++)
	{
	S1.run(ck);
	}
	}

[0422] The simulation is set up as before with queue objects and basic blocks. Next, a “sysgen” object is created, with name “S1”. All basic blocks in the simulation are appended to the “sysgen” objects by means of the $<<$ operator. If a timed basic block is to be used, as for instance in case of the “add” object, then the “fsm( )” pointer must be presented to “sysgen” rather then the basic block itself. A “sysgen” object knows how to run and combine both timed and untimed objects. For the description shown above, untimed versions of the file sources and sink “src” and “snk” will be used, while the timed version of the “add” object will be used.

[0423] Next, three clock cycles of the system are run. This is done by means of the “run(ck)” member function call of “sysgen”. The clock object “ck” is, because this simulation contains no registered signals, a dummy object. When running the simulator executable with stimuli file contents

[0424]

	SRC1	SRC2	-- not present in the file
	----	----	-- not present in the file
	1	4
	2	5
	3	6

you see the following appearing on the screen.

• [0425] *** INFO: Defining block SRC1
• [0426] *** INFO: Defining block SRC2
• [0427] *** INFO: Defining block ADD
• [0428] *** INFO: Defining block SNK1

[0429]

	fsm fsm: transition from go to go
	add#0
	add#1
	in	i1	1
	in	i2	4
	sig	ot	5
	out'	ot	5
	fsm fsm: transition from go to go
	add#0
	add#1
	in	i1	2
	in	i2	5
	sig	ot	7
	out'	ot	7
	fsm fsm: transition from go to go
	add#0
	add#1
	in	i1	3
	in	i2	6
	sig	ot	9
	out'	ot	9

[0430] The debugging output produced is enabled by the “setinfo( )” call on the “sysgen” object. The parameter “verbose” enables full debugging information. For each clock cycle, each fsm responds which transition it takes. The fsm of the “add” block is called “fsm”, an as is seen it makes transitions from the single state “go” to the obvious destination. Each signal flowgraph during this simulation is executed in two phases (below it is indicated why). During simulation, the value of each signal is printed.

[0431] Selecting the Simulation Verbosity

[0432] The “setinfo” member function call of “sysgen” selects the amount of debugging information that is produced during simulation. Four values are available:

• • [0433] “silent” will cause no output at all. This can significantly speed up your simulation, especially for large systems containing several hundred of signal flowgraphs.
  • [0434] “terse” will only print the transitions that fsm's make.
  • [0435] “verbose” will print detailed information on all signal updates.
  • [0436] “regcontents” will print a list the values of registered signals that change during the current simulation. This is by far the most interesting option if you are debugging at the system level: when nothing happens, for instance when all your timed descriptions are in some “hold” mode, then no ouput is produced. When there is a lot of activity, then you will be able to track all registered signals that change.

[0437] This example is part of a simulation containing 484 registerd signals and 483 signal flowgraphs. Using “setinfo(verbose)” here might require a good text editor to see what is happening—if anything will happen before your quota is exceeded.

For instance, the code fragment

[0438]

	sysgen S(‘‘S’’);
	S.setinfo(regcontents);
	int cycle;
	for (cycle=0; cycle < 100; cycle++)
	{
	cout << ‘‘> Cycle ’’ << cycle << ‘‘\n’’;
	S.run(ck);
	}

can produce an output as shown below.

[0439]

	> Cycle 18
		coef_ram_ir_2	0	1
		copy_step_flag	1	0
		ext_ready_out	1	0
		pc	15	16
		step_flag	1	0
	> Cycle 19
		coef_ram_ir_2	1	0
		coef_wr_adr	12	13
		hold_pc	0	16
		pc	16	17
		pc_ctl_ir_1	1	0
	> Cycle 20
		step_clock	0	1
	> Cycle 21
		copy_step_flag	0	1
		prev_step_clock	0	1
		step_flag	0	1

[0440] Three Phases are Better

[0441] Although you will be saved from the details behind two-phase simulation, it is worthwhile to see the motivation behind it.

[0442] When you run an “sfg” “by hand” using the “run( )” method of an “sfg”, the simulation proceeds in one phase: read inputs, calculate, produce ouput. The “sysgen” object, on the other hand, uses a two-phase simulation mechanism.

[0443] The origin is the following. In the presence of feedback loops, your system data flow simulation will need initial values on the communication queues in order to start the simulation. However, the code generator assumes the communication queues will translate to wiring. Therefore, there will never be storage in the implementation of a communication queue to hold these intitial values. OCAPI works around this by producing these initial values at runtime. This gives rise to a three-phase simulation: in the first phase, initial values are produced, while in the second phase, they are consumed again. This process repeats every clock cycle.

[0444] The three-phase simulation mechanism is also able to detect combinatorial loops at the system level. If there exists such a loop, then the first phase of the simulation will not produce any initial value on the system interconnect. Consequently, in the last phase there will be at least one signal flowgraph that will not be able to complete execution in the current clock cycle. In that case, OCAPI will stop the simulation. Also, you get a list of all signal flowgraphs that have not completed the current clock cycle, in addition to the queue statistics that are attached to these signal flowgraphs.

Hardware Code Generation

[0445] OCAPI allows you to translate all timed descriptions to a synthesizable hardware description.

• • [0446] For each timed description, you get a datapath “.dsfg” file, that can be entered into the Cathedral-3 datapath synthesis environment, converted to VHDL and postprocessed by Synopsys-dc logic synthesis.
  • [0447] For each timed description, you also get a controller “.dsfg” file, which is synthesized through the same environment.
  • [0448] You also get a glue cell, that interconnects the resulting datapath and controller VHDL file.
  • [0449] You get a system interconnect file, that integrates all glue cells in your system. For this system interconnect file, you optionally can specify system inputs and outputs, scan chain interconnects, and RAM interconnects. The file is VHDL.
  • [0450] Finally, you also get debug information files, that summarize the behavior of and ports on each processor.

[0451] Untimed blocks are not translated to hardware. The use of the actual synthesis environments will not be discussed in this section. It is assumed to be known by a person skilled in the art.

[0452] The generate( ) Call

[0453] The member call “generate( )” performs the code generation for you. In the adder example, you just have to add

• [0454] S1.generate( );
  at the end of the main function. If you would compile this description, and run it, then you would see things are not quite OK:
• [0455] *** INFO: Generating Systen Link Cell
• [0456] *** INFO: Component generation for S1
• [0457] *** INFO: C++ currently defines 5 sig, 4 _sig, 1 sfg.
• [0458] *** INFO: Generating FSMD fsm
• [0459] *** INFO: FSMD fsm defines 1 instructions
• [0460] DSFGgen: signal i1 has no wordlength spec.
• [0461] DSFGgen: signal i2 has no wordlength spec.
• [0462] DSFGgen: signal ot has no wordlength spec.
• [0463] DSFGgen: not all signals were quantized. Aborting.
• [0464] *** INFO: Auto-cleanup of sfg

[0465] Indeed, in the adder example up to now, nothing has been entered regarding wordlengths. During code generation, OCAPI does quite some consistency checking. The general advice in case of warnings and errors is: If you see an error or warning message, investigate it. When you synthesize code that showed a warning or error during generation, you will likely fail in the synthesis process too.

[0466] The “add” description is now extended with wordlengths. 8 bit wordlengths are chosen. You modify the “add” class to include the following changes.

[0467]

	void add::define( )
	{
	dfix wl(0,8,0);
	_sig i1(‘‘i1’’,wl);
	_sig i2(‘‘i2’’,wl);
	_sig ot(‘‘ot’’,wl) ;
	. . .
	}

[0468] After recompiling and rerunning the OCAPI program, you now see:

• [0469] *** INFO: Generating Systen Link Cell
• [0470] *** INFO: Component generation for S1
• [0471] *** INFO: C++ currently defines 5 sig, 4 _sig, 1 sfg.
• [0472] *** INFO: Generating FSMD fsm
• [0473] *** INFO: FSMD fsm defines 1 instructions
• [0474] *** INFO: C++ currently defines 31 sig, 21 _sig, 3 sfg.
• [0475] *** INFO: Auto-cleanup of sfg

[0476] In the directory where you ran this, you will find the following files:

• • [0477] “fsm_dp.dsfg”, the datapath description of “add”
  • [0478] “fsm_fsm.dsfg”, the controller description of “add”
  • [0479] “fsm.vhd”, the glue cell description of add
  • [0480] “S1.vhd”, the system interconnect cell
  • [0481] “fsm.ports”, a list of the I/O ports of “add”.

[0482] The glue cell “fsm.vhd” has the following contents (only the entity declaration part is shown).

[0483] Cath3 Processor for FSMD design fsm

[0484]

	library IEEE;
	use IEEE.std_logic_1164.all;
	entity fsm is
	port (
	reset: in std_logic;
	clk: in std_logic;
	i1: in std_logic_vector(7 downto 0);
	i2: in std_logic_vector(7 downto 0);
	ot: out std_logic_vector(7 downto 0)
	);
	end fsm;

[0485] Each processor has a reset pin, a clock pin, and a number of I/O ports, depending on the inputs and ouputs defined in the signal flowgraphs contained in this processor. All signals are mapped to “std_logic” or “std_logic_vector”. The reset pin is used for synchronous reset of the embedded finite state machine. If you need to initialize registered signals in the datapath, then you have to describe this explicitly in a signal flowgraph, and execute this upon the first transition out of the initial state.

[0486] The “fsm.ports” file, indicates which ports are read in in each transition. In the example of the “add” class, there is only one transition, which results in the following “.ports” file

[0487]

********** SFG fsmgogo0 **********
	Port #	I/O	Port	Q
	1	I	i1	i1
	2	I	i2	i2
	1	O	ot	o1

[0488] The name of an input or output signal is used as a port name, while the name of the queue associated to it relates to the system net name that will be connected to this port.

[0489] System Cell Refinements

[0490] The system link cell incorporates all glue cells of your current timed system description. These glue cells are connected if they read/write from the same system queue. There are some refinements possible on the “sysgen” object that will also allow you to indicate system level inputs and ouputs, scan chains, and RAM connections.

[0491] System inputs and ouputs are indicated with the “inpad( )” and “outpad( )” member calls of “sysgen”. In the example, this is specified as

[0492]

	. . .
	sysgen S1 (‘‘S1’’);
	dfix b8 (0,8,0);
	S1.inpad(i1, b8);
	S1.inpad (i2, b8);
	S1.outpad (o1, b8);

[0493] Making these connections will make the “i1”, “i2”, “o1” signals appear in the entity declaration of the system cell “S1”. The entity declaration inside of the file “S1.vhd” thus looks like

[0494]

	entity S1 is
	port (
	reset:	in std_logic;
	clk: in std_logic;
	i1: in std_logic_vector(7 downto 0);
	i2: in std_logic_vector(7 downto 0);
	o1: out std_logic_vector(7 downto 0)
	);
	end S1;

[0495] Scan chains can be added at the system level, too. For each scan chain you must indicate which processors it should include. Suppose you have three basic blocks (including a timed description and registers) with names “BLOCK1”, “BLOCK2”, “BLOCK3”. You attach the blocks to two scan chains using the following code.

[0496]

	scanchain SCAN1(“scan1”);
	scanchain SCAN2(“scan2”);
	SCAN1.addscan(& BLOCK1. fsm( ));
	SCAN1.addscan(& BLOCK2. fsm( ));
	SCAN2.addscan(& BLOCK3. fsm( ));

[0497] The “sysgen” object identifies the required scan chain connections through the “fsm” objects that are assigned to it. In order to have reasonable circuit test times, you should not include more then 300 flip-flops in each scan chain. If you have a processor that contains more then 300 flip-flops, then you should use another scan chain connection strategy.

[0498] Finally, you can generate code for the standard untimed block RAM. There are two possible interconnection mechanisms: the first will include the untimed RAM blocks in “sysgen” as internal components of the system link cell. The second will include the RAM blocks as external components. This latter method requires you to construct a new “system-system link cell”, that includes the RAM entities and the system link cell in a larger structure. However, it might be required in case you have to remap the standard RAM interface, or introduce additional asynchronous timing logic.

[0499] An example of the two methods is shown next

[0500]

	ram RAM1(“ram1”, addr1, di1, do1, wr, rd, 128);
	ram RAM2(“ram2”, addr2, di2, do2, wr, rd, 128);
	// types of address and data bus
	dfix addrtype(0, 7, 0);
	dfix dattype(0, 4, 0);
	sysgen S1 (‘‘S1’’);
	// define an external ram
	S1.extern_ram(RAM1, addrtype, dattype);
	// define an internal ram
	S1.intern_ram(RAM2, addrtype, dattype);

[0501] Pitfalls for Code Generation

[0502] As always, there are a number of pitfalls when things get complex. You should watch the following when diving into code generation.

[0503] OCAPI generates nicely formatted code, that you can investigate. To help you in this process, also the actual signal names that you have specified are regenerated in the VHDL and DSFG code. This implies that you have to stay away from VHDL and DSFG keywords, or else you will get an error from either Cathedral-3 or Synopsys.

[0504] The mapping of the fixed point library to hardware is, in the present release, minimal. First of all, although registered signals allow you to specify an initial value, you cannot rely on this for the hardware circuit. Registers, when powered on, take on a random state. Therefore, make sure that you specify the initialization sequence of your datapath. A second fixed point pitfall is that the hardware support for the different quantization schemes is lacking. It is assumed that you finally will use truncated quantization on the lsb-side and wrap-around quantization on the msb-side of all signals. The other quantization schemes require additional hardware to be included. If you really need, for instance, saturated msb quantization, then you will have to describe it in terms of the default quantization.

[0505] Finally, the current set of hardware operators in Cathedral-3 is designed for signed representations. They work with unsigned representations also as long as you do no use relational operations (<, > and the like). In this last case, you should implement the unsigned operation as a signed one with one extra bit.

Verification and Testbenches

[0506] Once you have obtained a gate level implementation of your circuit, it is necessary to verify the synthesis result. OCAPI helps you with this by generating testbenches and testbench stimuli for you while you run timed simulations and do code generations.

[0507] The example of the “add” class introduced previously is picked up again, and testbench generation capability is included to the OCAPI description.

[0508] Generation of Testbench Vectors

[0509] The next example performs a three cycle simulation of the “add” class and generates a testbench vectors for it.

[0510]

	#include “qlib.h”
	void main( )
	{
	dfbfix i1(“i1”);
	dfbfix i2(“i2”);
	dfbfix o1(“o1”);
	src SRC1(“SRC1”, i1, “SRC1”);
	src SRC2(“SRC2”, i2, “SRC2”);
	add ADD(“ADD”, i1, i2, o1);
	snk SNK1(“SNK1”, o1, “SNK1”);
	sysgen S1(“S1”);
	S1 << SRC1;
	S1 << SRC2;
	S1 << ADD.fsm( );
	S1 << SNK1;
	ADD.fsm( ).tb_enable( );
	clk ck;
	int i;
	for (i=0; i<3; i++)
	S1.run(ck);
	ADD.fsm( ).tb_data( );
	}

[0511] Just before the timed simulation starts, you enable the generation of testbench vectors by means of a “tb_enable( )” member call for each fsm that requires testbench vectors.

[0512] During simulation, the values on the input and ouput ports of the “add” processor are recorded. After the simulation is done, the testbenches are generated using a “tb\_data( )” member function call.

[0513] Testbench generation leaves three data files behind:

• • [0514] “fsm_tb.dat” contains binary vectors of all inputs of the “add” processor. It is intended to be read in by the VHDL simulator as stimuli.
  • [0515] “fsm_tb.dat_hex” contains hexadecimal vectors of all inputs and outputs of the “add” processor. It contains the output that should be produced by the VHDL simulator when the synthesis was successful.
  • [0516] “fsm_tb.dat_info” documents the contents of the stimuli files by saying which stimuli vector corresponds to which signal

[0517] When compiling and running this OCAPI program, the following appears on screen.

• [0518] *** INFO: Defining block SRC1
• [0519] *** INFO: Defining block SRC2
• [0520] *** INFO: Defining block ADD
• [0521] *** INFO: Defining block SNK1
• [0522] *** INFO: Creating stimuli monitor for testbench of FSMD fsm
• [0523] *** INFO: Generating stimuli data file for testbench fsm_tb.
• [0524] *** INFO: Testbench fsm_tb has 3 vectors.

[0525] Afterwards, you can take a look at each of the three generated testbenches.

[0526]

	-- file: fsm_tb.dat
	00000001 00000100
	00000010 00000101
	00000011 00000110
	-- file: fsm_tb.dat_hex
	01 04 05
	02 05 07
	03 06 09
	-- file: fsm_tb.dat_info
	Stimuli for fsm_tb contains 3 vectors for
	i1_stim	read
	i2_stim	read

[0527] Next columns occur only in _hex.dat file and are outputs

• • [0528] o1_stim write

[0529] You can now use the vectors in the simulator. But first, you must also generate a testbench driver in VHDL.

[0530] Generation of Testbench Drivers

[0531] To generate a testbench driver, simply call the “tb_enable( )” member function of the “add” fsm before you initiate code generation. You will end up with a VHDL file “fsm_tb.vhd” that contains the following driver.

[0532] Test Bench for FSMD Design fsm

[0533]

	library IEEE;
	use IEEE.std_logic_1164.all;
	use IEEE.std_logic_textio.all;
	use std.textio.all;
	library clock;
	use clock.clock.all;
	entity fsm_tb is
	end fsm_tb;
	architecture rtl of fsm_tb is
	signal reset: std_logic;
	signal clk:std_logic;
	signal i1: std_logic_vector(7 downto 0);
	signal i2: std_logic_vector(7 downto 0);
	signal ot: std_logic_vector(7 downto 0);
	component fsm
	port (
	reset: in std_logic;
	clk: in std_logic;
	i1: in std_logic_vector(7 downto 0);
	i2: in std_logic_vector(7 downto 0);
	ot: out std_logic_vector(7 downto 0)
	);
	end component;
	begin
	crystal(clk, 50 ns);
	fsm_dut: fsm
	port map (
	reset => reset,
	clk => clk,
	i1 => i1,
	i2 => i2,
	ot => ot
	);
	ini: process
	begin
	reset <= ‘1’;
	wait until clk'event and clk = ‘1’;
	reset <= ‘0’;
	wait;
	end process;
	input:	process
	file stimuli : text is in “fsm_tb.dat”
	variable aline : line;
	file stimulo : text is out “fsm_tb.sim_out”;
	variable oline : line;
	variable v_i1: std_logic_vector(7 downto 0);
	variable v_i2: std_logic_vector(7 downto 0);
	variable v_ot: std_logic_vector(7 downto 0);
	variable v_i1_hx: std_logic_vector(7 downto 0);
	variable v_i2_hx: std_logic_vector(7 downto 0);
	variable v_ot_hx: std_logic_vector(7 downto 0);
	begin
	wait until reset'event and reset = ‘0’;
	loop
	if(not(endfile(stimuli))) then
	readline(stimuli, aline);
	read(aline, v_i1);
	read(aline, v_i2);
	else
	assert false
	report “End of input file reached”
	severity warning;
	end if;
	i1 <= v_i1;
	i2 <= v_i2;
	wait for 50 ns;
	v_ot := ot;
	v_i1_hx := v_i1;
	v_i2_hx := v_i2;
	v_ot_hx := v_ot;
	hwrite(oline, v_i1_hx);
	write(oline, ' ');
	hwrite(oline, v_i2_hx);
	write(oline, ' ');
	hwrite(oline, v_ot_hx);
	write(oline, ' ');
	writeline (stimulo, oline);
	wait until clk'event and clk = ‘1’;
	end loop;
	end process;
	end rtl;
	configuration tbc_rtl of fsm_tb is
	for rtl
	for all : fsm
	use entity work.fsm(structure);
	end for;
	end for;
	end tbc_rtl;

[0534] The testbench uses one additional library, “clock”, which contains the “crystal” component. This component is a simple clock generator that drives a 50% duty cycle clk.

[0535] This testbench will generate a file “fsm_tb.sim_out”. After running the testbench in VHDL, this file should be exactly the same as the “fsm_tb.dat_hex”. You can use the unix “diff” command to check this. The only possible differences can occur in the first few simulation cycles, if the VHDL simulator initializes the registers to “X”.

[0536] Using automatic testbench generation greatly speedups the verification process. You should consider using it whenever you are into code generation.

Compiled Code Simulations

[0537] For large designs, simulation speed can become prohibitive. The restricting factor of OCAPI is that the signal flowgraph data structures are interpreted at runtime. In addition, runtime quantization (fixed point simulation) takes up quite some CPU power.

[0538] OCAPI allows you to generate a dedicated C++ simulator, that runs compiled code instead of interpreted code. Also, additional optimizations are done on the fixed point simulation. The result is a simulator that runs one to two orders of magnitude faster then the interpreted OCAPI simulation. This speed increase adds up to the order of magnitude that interpreted OCAPI already gains over event-driven VHDL simulation.

[0539] As an example, a 75 Kgate design was found to run at 55 cycles per second (on a HP/9000). This corresponds to “4.1 million” gates per second, and motivates why C++ is the way to go for system synthesis.

[0540] Generating a Compiled Code Simulator

[0541] The compiled code generator is integrated into the “sysgen” object. There is one member function, “compiled( )”, that will generate this simulator for you.

[0542]

	#include ‘‘qlib.h’’
	#include ‘‘add.h’’
	void main( )
	{
	dfbfix i1(“i1”);
	dfbfix i2(“i2”);
	dfbfix o1(“o1”);
	add ADD(“ADD”, i1, i2, o1);
	sysgen S1(“S1”);
	S1 << ADD.fsm( );
	S1.compiled( );
	}

[0543] In this simple example, a compiled code generator is made for a design containing only one FSM. The generator allows to include several fsm blocks, in addition to untimed blocks.

[0544] When this program is compiled and run, it leaves behind a file “S1_ccs.cxx”, that contains the dedicated simulator. For the OCAPI user, the simulator defines one procedure, “one_cycle( )”, that simulates one cycle of the system.

[0545] When calling this procedure, it also produces debugging ouput similar to the “setinfo(regcontents)” call for “ctlfsm” objects. This procedure must be linked to a main program that will execute the simulation.

[0546] If an untimed block is present in the system, then it will be included in the dedicated simulator. In order to declare it, you must provide a member function “CCSdecl(ofstream &)” that generates the required C++ declaration. As an example, the basic RAM block declares itself as follows:

[0547]

	-- file: ram.h
	class ram : public base
	{
	public:
	. . .
	ram (char * name,
	FB& _address,
	FB& _data_in,
	FB& _data_out,
	FB& _w,
	FB& _r,
	int _size);
	void CCSdecl(ofstream &os;);
	. . .
	private:
	. . .
	};
	-- file: ram.cxx
	void ram::CCSdecl(ofstream &os;)
	{
	os << “ #include \”ram.h\“\n”;
	os << “ ram ” << typeName( ) << “(”;
	os << “\”“ << typeName( ) << “\”,″;
	os << address.name( ) << ″, ″;
	os << data_in.name( ) << ″,″;
	os << data_out.name( ) << ″,″;
	os << w.name( ) << ″,″;
	os << r.name( ) << ″,″;
	os << size << ″);\n″;
	}

[0548] This code enables the ram to reproduce the declaration by which it was originally constructed in the interpreted OCAPI program. Every untimed block that inherits from “base”, and that you whish to include in the compiled code simulator must use a similar “CCSdecl” function.

[0549] Compiling and Running a Compiled Code Simulator

[0550] The compiled code simulator is compiled and linked in the same way as a normal OCAPI program. You must however also provide a “main” function that drives this simulator.

[0551] The following code contains an example driver for the “add” compiled code simulator.

[0552]

	#include “qlib.h”
	void one_cycle( );
	extern FB i1;
	extern FB i2;
	extern FB o1;
	void main( )
	{
	i1 << dfix(1) << dfix(2) << dfix(3);
	i2 << dfix(4) << dfix(5) << dfix(6);
	one_cycle( );
	one_cycle( );
	one_cycle( );
	while(o1.getSize( ))
	cout << o1.get( ) << “\n”;
	}

[0553] When run, this program will produce the same results as before. In contrast to the compiled simulaton of your MPEG-4 image processor, you will not be able to notice any speed increase on this small example.

Faster Communications

[0554] OCAPI uses queues as a means to communicate during simulation. These queues however take up CPU power for queue management. To save this power, there is an additional queue type, “wireFB”, which is used for the simulation of point-to-point wiring connections.

[0555] The dfbfix_wire Class

[0556] A “wireFB” does not move data. In contrast, it is related to a registered driver signal. At any time, the value read of this queue is the value defined by the registered signal. Because of this signal requirement, a “wireFB” cannot be used for untimed simulations. The following example of an accumulator shows how you can use a “wireFB”, or the equivalent “dfbfix_wire”.

[0557]

	#include “qlib.h”
	void main( )
	{
	clk ck;
	_sig a(“a”,ck,dfix(0));
	_sig b(“b”);
	dfbfix_wire A(“A”,a);
	dfbfix B(“B”);
	sfg accu;
	accu.starts( );
	a = a + b;
	accu << “accu”;
	accu << ip(b, B);
	accu << op(a, A);
	accu.check( );
	B << dfix(1) << dfix(2) << dfix(3);
	while (B.getSize( ))
	{
	accu.eval(cout);
	accu.tick(ck);
	}
	{

[0558] A “wireFB” is identical in use as a normal “FB”}. Only, for each “wireFB”, you indicate a registered driver signal in the constructor.

[0559] Interconnect Strategies

[0560] The “wireFB” object is related to the interconnect strategy that you use in your system. An interconnect strategy includes a decision on bus-switching, bus-storage, and bus-arbitration. OCAPI does not solve this problem for you: it depends on your application what the right interconnection strategy is.

[0561] One default style of interconnection provided by OCAPI is the point-to-point, register driven bus scheme. This means that every bus carries only one signal from one processor to another. In addition, bus storage in included in the processor that drives the bus.

[0562] More complex interconnect strategies, like the one used in Cathedral-2, are also possible, but will have to be described in OCAPI explicitly. Thus, the freedom of target architecture is not without cost. In the section “Meta-code generation”, a solution to this specification problem is presented.

Meta-code Generation

[0563] OCAPI internally uses meta-code generation. With this, it is meant that there are code generators that generate new “fsm”, “sfg” and “sig” objects which in turn can be translated to synthesizable code.

[0564] Meta-code generation is a powerful method to increase the abstraction level by which a specification can be made. This way, it is also possible to make parametrized descriptions, eventually using conditions. Therefore, it is the key method of soft-chip components, which are software programs that translate themselves to a wide range of implementations, depending on the user requirements.

[0565] The meta-code generation mechanism is also available to the user. To demonstrate this, a class will be presented that generates an ASIP datapath decoder.

[0566] An ASIP Datapath Idiom

[0567] An ASIP datapath, when described as a timed description within OCAPI, will consist of a number of signal flowgraphs and a finite state machine. The signal flowgraphs express the different functions to be executed by the datapath. The fsm description is a degenerated one, that will use one transition per decoded instruction. The transition condition is expressed by the “instruction” input, and selects the appropriate signal flowgraph for execution.

[0568] Because the finite state machine has a fixed, but parametrizable structure, it is subject for meta-code generation. You can construct a “decoder” object, that generates the “fsm” for you. This will allow compact specification of the instruction set.

[0569] First, the “decoder” object (which is present in OCAPI) itself is presented.

[0570]

	-- the include file
	#define MAXINS 100
	#include “qlib.h”
	class decoder : public base
	{
	public:
	decoder(char *_name, clk &ck;, dfbfix &_insq);
	void dec(int _numinstr);
	ctlfsm &fsm;( );
	void dec(int _code, sfg &);
	void dec(int _code, sfg &, sfg &);
	void dec(int _code, sfg &, sfg &, sfg &);
	private:
	char *name;
	clk *ck;
	dfbfix *insq;
	int inswidth;
	int numinstr;
	int codes [MAXINS];
	ctlfsm _fsm;
	state active;
	sfg decode;
	_sigarray *ir;
	cnd * deccnd(int);
	void decchk(int);
	};
	-- the .cxx file
	#include “decoder.h”
	static int numbits(int w)
	{
	int bits = 0;
	while (w)
	{
	bits++;
	w = w >> 1;
	}
	return bits;
	}
	int bitset (int bitnum, int n)
	{
	return (n & (1 << bitnum));
	}
	decoder::decoder(char *_name, clk &_ck;, dfbfix &_insq;)
	: base(_name)
	{
	name = _name;
	insq = _insq.asSource(this);
	ck = &_ck;
	numinstr = 0;
	inswidth = 0;
	_fsm << _name;
	// active << strapp(name, “_go_”);
	active << “go”;
	_fsm << deflt(active);
	}
	void decoder::dec(int n)
	{
	// define a decoder that decodes n instructions
	// instruction numbers are 0 to n−1
	// create also the instruction register
	if (! (n>0))
	{
	cerr << “*** ERROR: decoder “ << name << ” must
	have at least one instruction\n”;
	exit (0);
	}
	inswidth = numbits (n−1);
	if (n > MAXINS)
	{
	cerr << “*** ERROR: decoder “ << name << ”
	exceeds decoding capacity\n”;
	exit(0);
	}
	dfix bit (0,1,0,dfix::ns);
	ir = new _sigarray((char *) strapp(name, “_ir”),
	inswidth, ck, bit);
	decode.starts( );
	int i;
	SIGW(irw, dfix(0, inswidth, 0, dfix::ns));
	for (i=0; i<inswidth; i++)
	{
	if(i)
	(*ir)[i] = cast (bit, irw >>
	_sig(dfix(i, inswidth, 0,dfix::ns)));
	else
	(*ir)[i] = cast (bit, irw);
	}
	decode << strapp(“decod”, name);
	decode << ip(irw, *insq);
	}
	void decoder::decchk(int n)
	{
	// check if the decoder can decode this instruction
	int i;
	if (!inswidth)
	{
	cerr << “*** ERROR: decoder “ << name << ” must
	first define an instruction width\n”;
	exit (0);
	}
	if (n > ((1 << inswidth)−1))
	{
	cerr << “*** ERROR: decoder “ << name << ”
	cannot decode code “ << n << “\n”;
	exit(0);
	}
	for (i=0; i<numinstr; i++)
	{
	if (n == codes[i])
	{
	cerr << “*** ERROR: decoder ” << name << ”
	decodes code ” << n << “ twice\n”;
	exit (0);
	}
	}
	codes [numinstr] = n;
	numinstr++;
	}
	cnd *decoder::deccnd(int n)
	{
	// create the transition condition that corresponds
	// to the instruction number n
	int i;
	cnd *cresult = 0;
	if (bitset(0, n))
	cresult = &_cnd;((*ir)[0]);
	else
	cresult = &(!_cnd((*ir)[0]));
	for (i = 1; i < inswidth; i++)
	{
	if (bitset(i, n))
	cresult = &(*cresult && _cnd((*ir)[i]));
	else
	cresult = &(*cresult && !_cnd((*ir)[i]));
	}
	return cresult;
	}
	void decoder::dec(int n, sfg &s;)
	{
	// enter an instruction that executes one sfg
	decchk(n);
	active << *deccnd(n) << decode << s << active;
	}
	void decoder::dec(int n, sfg &s1;, sfg &s2;)
	{
	// enter an instruction that executes two sfgs
	decchk(n);
	active << *deccnd(n) << decode << s1 << s2 <<
	active;
	}
	void decoder::dec (int n, sfg &s1, sfg &s2, sfg &s3)
	{
	// enter an instruction that executes three sfgs
	decchk(n);
	active << *deccnd(n) << decode << s1 << s2 << s3 <<
	active;
	}
	ctlfsm & decoder::fsm( )
	{
	return _fsm;
	}

[0571] The main principles of generation are the following. Each instruction for the ASIP decoder is defined as a number, in addition to one to three signal flowgraphs that need to be executed when this instruction is decoded. The “decoder” object keeps track of the instruction numbers already used and warns you if you introduce a duplicate. When the instruction number is unique, it is split up into a number of instruction bits, and a fsm transition condition is constructed from these bits.

[0572] The ASIP Datapath at Work

[0573] The use of this object is quite simple. In a timed description were you want to use the decoder instead of a plain “fsm”, you inherit from this decoder object rather then from the “base” class. Next, instead of the fsm description, you give the instruction list and the required signal flowgraphs to execute.

[0574] As an example, an add/subtract ASIP datapath is defined. We select addition with instruction number 0, and subtraction with instruction number 1. The following code (that also uses the supermacros) shows the specification. The inheritance to “decoder” also establishes the connection to the instruction queue.

[0575]

	-- include file
	#ifndef ASIP_DP_H
	#define ASIP_DP_H
	class asip_dp : public decoder
	{
	public:
	asip_dp  (char *name,
	clk &ck,
	FB &ins,
	_PRT(in1),
	_PRT(in2),
	_PRT(o1));
	private:
	PRT(in1);
	PRT(in2);
	PRT(o1 );
	};
	-- code file
	#include “asip_dp.h”
	dfix typ(0,8,0);
	asip_dp::asip_dp  (char *name,
	clk &ck,
	FB &ins,
	_PRT(in1),
	_PRT(in2),
	_PRT(o1)) :  decoder(name, ck, ins),
	IS_SIG(in1, typ),
	IS_SIG(in2, typ),
	IS_SIG(o1, typ)
	{
	IS_IP(in1);
	IS_IP(in2);
	IS_OP(o1);
	SFG(add);
	GET(in1);
	GET(in2);
	o1 = in1 + in2;
	PUT(o1);
	SFG(sub);
	GET(in1);
	GET(in2);
	ol = in1 − in2;
	PUT(o1);
	dec(2); // decode two instructions
	dec(0, SFGID(add));
	dec(1, SFGID(sub));
	}

[0576] To conclude, one can note that meta-code generation allows reuse of design “idioms” (classes) rather then design “instances” (objects). Intellectual-property code generators are a direct consequence of this.

[0577] In the design of a telecommunication system (FIG. 1A), we distinguish four phases: link design, algorithm design, architecture design and circuit design. These phases are used to define and model the three key components of a communication system: a transmitter, a channel model, and a receiver.

• • [0578] The link design (1) is the requirement capture phase. Based on telecommunication properties such as transmission bandwidth, power, and data throughput (the link requirements), the system design space is explored using small subsystem simulations. The design space includes all algorithms which can be used by a transmitter/receiver pair to meet the link requirements. Out of receiver and transmitter algorithms with an identical functionality, those with minimal complexity are preferred. Besides this exploration, any expected transmission impairment must also be modeled into a software channel model.
  • [0579] The algorithm design (2) phase selects and interconnects the algorithms identified in the link design phase. The output is a software algorithmic description in C++ of digital transmitter and receiver parts in terms of floating point operations. To express parallelism in the transmitter and receiver algorithms, a data-flow data model is used. Also, the transmission imperfections introduced by analog parts such as the RF front-ends are annotated to the channel model.
  • [0580] The architecture design (3) refines the data model of the transmitter or receiver. The target architectural style is optimized for high speed execution, uses distributed control semantics and pipeline mechanisms. The resulting description is a fixed point, cycle true C++ description of the algorithms in terms of execution on bit-parallel operators. The architecture design is finished with a translation of this description to synthesizable VHDL.
  • [0581] Finally, circuit design (4) refines the bit-parallel implementation to circuit level, including technology binding, the introduction of test hardware, and design rule checks.
    Target Architecture

[0582] The target architecture (5), shown in FIG. 2, consists of a network of interconnected application specific processors. Each processor is made up of bit-parallel data-paths. When hardware sharing is applied, also a local control component is needed to perform instruction sequencing. The processors are obtained by behavioral synthesis tools or RT level synthesis tools. In either case, circuits with a low amount of hardware sharing are targeted. The network is steered by one or multiple clocks. Each clock signal defines a clock region. Inside a clock region the phase relations between all register clocks are manifest. Clock division circuits are used to derive the appropriate clock for each processor.

[0583] In between each processor, a hardware queue is present to transport data signals. They increase parallelism inside a clock region and maintain consistency between different streams of data arriving at one processor.

[0584] Across clock region boundaries, synchronization interfaces are used. These interfaces detect the presence of data at the clock region boundary and gate clock signals for the clock region that they feed. This way, non-manifest and variable data rates in between clock regions are supported.

[0585] The ensemble of clock dividers and handshake circuits forms a parallel scheduler in hardware, synchronizing the processes running on the bit-parallel processor.

Overview of the C++ Modeling Levels

[0586] An overview of the distinct C++ modeling levels used by OCAPI is given in FIG. 3. The C++ modeling spans three subsequent levels in the design flow: the link level, the algorithm level and the architecture level. The transition to the last level, the circuit level, is made by automated means trough code generation. Usually, VHDL is used as the design language in this lowest level.

[0587] The link level is available through data-vector modeling. Using a design mechanism called parallelism scaling, this level is refined to the algorithm level. The algorithm level uses data-flow semantics. Using two distinct refining mechanisms in the data-flow level, we can refine this level to a register transfer level.

[0588] The two refining mechanisms are clock cycle true modeling and fixed point modeling. Clock cycle true modeling is achieved by allocating cycle budgets and operators for each algorithm. To help the designer in this decision, operation profiling is foreseen. Fixed point modeling restricts the dynamic range of variables in the algorithms to a range for which a hardware operator can be devised. Signal statistics are returned by the design to help the designer with this.

[0589] The last level, the architecture model, uses a signal flowgraph to provide a behavioral description. Using this description synthesizable code is generated. The resulting code then can be mapped onto gates using a register-transfer design tool such as DC of Synopsys.

Data-vector Modeling

[0590] The upper level of representation of a communication system is the link level. It has the following properties:

• • [0591] It uses pure mathematical manipulation of functions. Time is explicitly manipulated and results in irregular-flow descriptions.
  • [0592] It uses abstraction of all telecommunication aspects that are not relevant to the problem at hand.

[0593] In this representation level, MATLAB is used for simulation. MATLAB uses the data-vector as the basic data object. To represent time functions in MATLAB, they are sampled at an appropriate rate. Time is present as one of the many vector dimensions. For example, the MATLAB vector addition
a=b+c;
can mean both sequential addition in time (if the b and c vectors are thought of as time-sequential), or parallel addition (if b and c happen to be defined at one moment in time). MATLAB simply make no distinction between these two cases.

[0594] Besides this time-space feature, MATLAB has a lot of other properties that makes it the tool-of-choice within this design level:

• • [0595] The ease with which irregular flow of data is expressed with vector operations. For example, the operation max(vector), or std(vector).
  • [0596] The flexibility of operations. A maximum operation on a vector of 10 elements or 1000 elements looks identically: max(vector).
  • [0597] The interactivity of the tool, and the transparency of data object management.
  • [0598] The extended library of operations, that allow very dense description of functionality.
  • [0599] Graphics and simulation speed.

[0600] This data-vector restriction is to be refined to a data-flow graph representation of the system. Definition of the data-flow graph requires definition of all actors in the graph (actor contents as well as actor firing rules) and definition of the graph layout.

[0601] In order to design systems effectively with the SOC++ design flow, a smooth transition between the data-vector level and the data-flow level is needed. A script to perform this task is constructed as can be seen in the following example.

Initial Data-Vector Description

[0602] We consider a pseudonoise (PN) code correlator inside a direct sequence spread-spectrum (DS/SS) modem as an example (FIG. 4).

[0603]

	% input data
	in = [1 2 1 3 3 4 1 2] ;
	% spreading code
	c = [1 −1 1 −1] ;
	% correlate
	ot = corr (in, c)
	% find correlation peak
	[max, maxpos] = max (ot) ;

[0604] A vector of input data in is defined containing 8 elements. These are subsequent samples taken from the chip demodulator in the spread spectrum modem. The dimension of in thus corresponds to the time dimension. The input vector in is in principle infinite in length. For simulation purposes, it is restricted to a data set which has the same average properties (distribution) as the expected received data.

[0605] The samples of in are correlated with the PN-code vector of length 4, c. The output vector ot thus contains 5 samples, corresponding to the five positions of in at which c can be aligned to. The max function locates the maximum value and position inside the correlated data. The position maxpos is subsequently used to synchronize the PN-code vector with the incoming data and thus is the desired output value of the algorithm.

[0606] This code is an elegant and compact specification, yet it offers some open questions for the PN-correlator designer:

• • [0607] The algorithm has an implicit startup-effect. The first correlation value can only be evaluated after 4 input samples are available. From then on, each input sample yields an additional correlation value.
  • [0608] The algorithm misses the common algorithmic iteration found in digital signal processing applications: each statement is executed only once.
  • [0609] For the implementation, no statement is made regarding the available cycle budget. This is however an important specification for the attainable acquisition speed of the modem.

[0610] All of these questions are caused by the parallelism of the data-vector description.

[0611] We now propose a way to make the parallelism of the operations more visible. Each of the MATLAB operations is easily interpreted. Inside the MATLAB simulation, the length of the operands will first be determined in order to select the correct operation behavior. For example,

• [0612] [max, maxpos]=max(ot)
  determines the maximum on a vector of length 5 (which is the length of the operand ot). It needs at least 4 scalar comparisons to evaluate the result. If ot would for example have a longer length, more scalar comparisons would be needed. To indicate this in the description, we explicitly annotate each specific instance of the generic operations with the length of the input vectors.

[0613]

	% input data
	in =  [1 2 1 3 3 4 1 2] ;
	8
	% spreading code
	c =  [1 −1 1 −1] ;
	4
	% correlate
	ot  =  corr  (in, c)
	5   8,4
	% find correlation peak
	[max, maxpos] =  max  (ot) ;
	1  5

[0614] This little annotation helps us to see the complexity of the operations more clearly. We will use this when considering implementation of the description in hardware. It is of course not the intention to force a user to do this (MATLAB does this already for him/her).

[0615] When thinking about the implementation of this correlator, one can imagine different realizations each having a different amount of parallelism, that is, the mapping of all the operations inside corr( ) and max( ) onto a time/space axis. This is the topic of the next section.

Scaled Description

[0616] Consider again the definition of the PN code, as in:

[0617]

	% spreading code
	c =  [1 −1 1 −1] ;
	4

[0618] This MATLAB description defines the variable c to be a data-vector containing 4 different values. This vector assignment corresponds to 4 concurrent scalar assignments. We therefore say that the maximal attainable parallelism in this statement is 4.

[0619] In order to achieve this parallelism in the implementation, there must be hardware available to perform 4 concurrent scalar assignments. Since a scalar assignment in hardware corresponds to driving a data bus to a certain state, we need 4 busses in the maximal parallel implementation. If only one bus would be desired, then we would have to indicate this. For each of the statements inside the MATLAB description, a similar story can be constructed. The indication of the amount of parallelism is an essential step in the transition from data-vectors to data-flow. We call this the scaling of parallelism. It involves a restriction of the unspecified communication bandwidth in the MATLAB description to a fixed number of communication busses. It is indicated as follows in the MATLAB description.

[0620]

	% input data
	in =  [1 2 1 3 3 4 1 2] ;
	8@1
	% spreading code
	c =  [1 −1 1 −1] ;
	4@4
	% correlate
	ot  =  corr  (in, c)
	5@1  8,4
	% find correlation peak
	[max, maxpos] =  max  (ot) ;
	1@1  5

[0621] As is seen, each assignment is extended with a @i annotation, that indicates how the parallelism in the data vectors is ordened onto a time axis. For example, the 8 input values inside in are provided sequentially by writing 8@1. The 4 values of c on the other hand, are provided concurrently. We see that, whatever implementation of the corr operation we might use, at least 8 iterations will be required, simply to provide the data to the operation.

[0622] At this moment, the description is getting closer to the data-flow level, that uses explicit iteration. One more step is required to get to the data flow graph level. This is the topic of the next section.

Data Flow Graph Definition

[0623] In order to obtain a graph, the actors and edges inside this graph must be defined. Inside the annotated MATLAB description, data precedences are already present through the presence of the names of the vectors. The only thing that is missing is the definition of actor boundaries; edges will then be defined automatically by the data precedences going across the actor boundaries.

[0624] This can be done by a new annotation to the MATLAB description. Three actors will be defined in the DS/SS correlator.

[0625]

	actor1 {
	% input data
	in =  [1 2 1 3 3 4 1 2] ;
	8@1
	}
	actor2 {
	% spreading code
	c =  [1 −1 1 −1] ;
	4@4
	% correlate
	ot  =  corr  (in, c)
	5@1   8,4
	}
	actor3 {
	% find correlation peak
	[max, maxpos] =   max   (ot) ;
	1@1   5
	}

[0626] Again the annotation should be seen as purely conceptual; it is not intended for the user to write this code. Given these annotations, a data flow graph can be extracted from the scaled MATLAB description in an unambiguous way.

• • [0627] actor1 is an actor with no input, and one output, called in.
  • [0628] actor2 is an actor with 1 input in and one output ot.
  • [0629] actor3 is an actor with 1 input ot and outputs maxpos and max.

[0630] Furthermore, the simulation uses queues to transport signals in between the actors. We need three queues, called in, ot and maxpos.

[0631] The missing piece of information for simulation of this dataflow graph are the firing rules (or equivalently the definition of productions and consumptions on each edge). A naive data flow model is shown in FIG. 4: actor1 (10) produces 8 values, which are correlated by actor2 (11), while the maximum is selected inside actor3 (12).

[0632] This would however mask the parallelism scaling operation inside the MATLAB description. For example, it was chosen to provide the 8 values of the in vector in a sequential way over a parallel bus. It is believed that the multi-rate SDF model therefore is not a good container for the annotated MATLAB description.

[0633] Another approach is a cyclostatic description. In this case we have a graph as in FIG. 5.

[0634] We see that the determination of production patterns involves examining the latencies of operations internal to the actor. This increases the complexity of the design script. It is simpler to perform a demand driven scheduling of all actors. The firing rule only has to examine the availability of input tokens.

[0635] The desired dataflow format as in FIG. 6 is thus situated in between the multirate SDF level and the cyclostatic SDF level. It is proposed to annotate consumptions and productions in the same way as it was written down in the matlab description:

• • [0636] 8@1 is the production of actor1. It means: 8 samples are produced one at a time.
  • [0637] 8@1 and 5@1 is the consumption and production of actor2 respectively.
  • [0638] 5@1 and 1@1, 1@1 are the consumption and productions for actor3.
    Data-flow Simulation

[0639] Given an annotated matlab description, a simulation can now be constructed by writing a high-level model for each actor, interconnecting these with queues and constructing a system schedule. OCAPI provides both a static scheduler and a demand-driven scheduler.

[0640] Out of this simulation, several statistics are gathered:

• • [0641] On each queue, put and get counts are observed, as well as signal statistics (minimum and maximum values). The signal statistics provide an idea of the required buswidths of communication busses.
  • [0642] The scheduler counts the firings per actor, and operation executions (+, −, *, . . . ) per actor. This profiling helps the designer in deciding cycle budgets and hardware operator allocation for each actor.

[0643] These statistics are gathered through a C++ operator overloading mechanism, so the designer gets them for free if he uses the appropriate C++ objects (schedule, queue and token class types) for simulation.

[0644] We are next interested in the detailed clock-cycle true behavior of the actors and the required storage and handshake protocol circuits on the communication busses. This is the topic of the next step, the actor definition.

Actor Definition

[0645] The actor definition is based on two elements:

• • [0646] Signal-flowgraph representation of behavior.
  • [0647] Time-verification of the system.

[0648] The two problems can be solved independently using the annotated MATLAB code as specification. In OCAPI:

• • [0649] The actor RT modeling proceeds in C++ and can be freely intermixed with high level descriptions regarding both operator wordlength effects and clock-cycle true timing.
  • [0650] The time-verification approach allows the system feasibility to be checked at all times by warning the designer for deadlock and/or causality violations of the communication.
    Signal Flowgraph Definition

[0651] Within the OCAPI design flow, a class library was developed to simulate behavior at RT-level. It allows

• • [0652] To express the behavior of an algorithm with arbitrary implementation parallelism by setting up an signal flow graph (SFG) data structure.
  • [0653] To simulate the behavior of an actor at a clock-cycle true level by interpreting this SFG data structure with instantiated token values.
  • [0654] To specify wordlength characteristics of operations regarding sign, overflow and rounding behavior. Through explicit modeling of the quantization characteristic rather than the bit-vector representation (as in SPW), efficient simulation runtimes are obtained.
  • [0655] To generate C++ code for this actor, and hence perform the clock cycle true simulation with compiled code.
  • [0656] To generate VHDL code for this actor, and synthesize an implementation with Synopsys DC.
  • [0657] To generate DSFG code for this actor, and synthesize an implementation with Cathedral-3. It was observed that Cathedral-3 performs a better job with relation to both critical path and area of the obtained circuits than Synopsys DC. The best synthesis results are obtained by first using Cathedral-3 to generate a circuit at gate level and then Synopsys-DC to perform additional logic optimization as a postprocessing.

[0658] An important observation was made regarding simulation speed. For equivalent descriptions at different granularities, the following relative runtimes were found:

• • [0659] 1 for the MATLAB simulation.
  • [0660] 2 for the untimed, high level C++ data flow description.
  • [0661] 4 for the timed, fixed point C++ description (compiled code).
  • [0662] 40 for the procedural, word-level VHDL description.

[0663] It is thus concluded that RT-modeling of systems within OCAPI is possible within half an order of magnitude of the highest level of description. VHDL modeling however, is much slower. Currently the figure of 40 times MATLAB is even considered an under-estimate. Future clock-cycle based VHDL simulators can only solve half of this problem, since they still use bit-vector based simulation of tokens rather then quantization based simulation.

[0664] Next, the modeling issues in C++ are shown in more detail. The C++ signal-flowgraph representation uses a signal data-type, that can be either a registered or else an immediate value. With this data-type, expressions are formed using the conventional scalar operations. (+, −, *, shifts and logical operations). Expressions are grouped together in a signal flowgraph. A signal flowgraph interfaces with the system through the data-flow simulation queues. Several signal-flowgraphs can be grouped together to a SFG-sequence. A SFG sequence is an expression of behavior that spans several cycles. The specification is done through a finite state machine model, for which transition conditions can be expressed. The concept of SFG modeling is pictured in FIG. 7.

[0665] The combination of different SFG's in combination with a finite state machine make up the clock-cycle true actor model. Within the actor, SFG communication proceeds through registered signals. Communication over the boundaries of an actor proceeds through simulation queues.

[0666] When the actor is specified in this way, and all signal wordlengths are annotated to the description, an automated path to synthesis is available. Several different SFG's can be assigned to one datapath. Synthesizable code is generated in such a way that hardware sharing between different sfg's is possible. A finite state machine (FSM) description is first translated to SFG format to generate synthesizable code in the same way. There is an implicit hierarchy available with this method: by assigning different FSM-SFG's to one datapath, an overall processor architecture is obtained that again has a mode port and therefore looks like a (multicycle) datapath. For macro control problems (such as acquisition/tracking algorithm switching in modems), this is a necessity.

[0667] Although the distance between the annotated MATLAB level and this RT-level SFG seems large, it is reasonable on the actor level. Consider for example

[0668]

	actor3 {
	% find correlation peak
	[max, maxpos] =  max  (ot) ;
	1@1   5
	}

[0669] We are asked here to write time the max( ) operation with an SFG. actor2 has scaled the parallelism of ot to 5@1. A solution is presented in actual C++ code.

[0670]

	{
	FB qin(“qin”) ;	//input queue
	FB q1out(“qout”) ;	//output queue
	FB q2out(“qout”) ;	//output queue
	FB start(“start”) ;	//the start pin of the
		processor
	clock ck ;
	_sig currmax(ck,dfix(0)) ;	//registry holding current
		maximum
	_sig maxpos(ck,dfix(0)) ;	//registry holding position
		of max
	_sig currpos(ck,dfix(0)) ;	// current position
	_sig inputvalue ;	//holds input values
	_sig maxout ;
	_sig maxposout ;
	_sig one(dfix(1)) ;	//a constant
	SFG sfg0, sfg1,sfg2 ;	//we use 3 sfg's
	sfg0.starts( ) ;	//code after this is for sfg0
	currmax = inputvalue ;
	maxpos = one ;
	currpos = one ;
		//next, give sfg0 a mode and
		an input queue
	sfg0 <<“m0”<<ip(inputvalue,qin) ;
	sfg1.starts( ) ;	//code after this is for sfg1
		//this is a conditional
		assignment
	currmax=(inputvalue>currmax).cassign (inputvalue, currmax);
	maxpos = (inputvalue > currmax).cassign(currpos, maxpos) ;
	currpos = currpos + 1 ;
	sfg1 <<“m1”<<ip(inputvalue,qin) ;
	sfg2.starts( ) ;	//the last SFG
	maxposout=(inputvalue>currmax).cassign (_sig (dfix(4)),maxpos);
	maxout=(inputvalue>currmax).cassign (inputvalue, currmax) ;
	sfg2 <<“m2”<< op(maxout,qout) << op(maxposout,q2out) ;
	state s0(“s0”), s1(“s1”), s2(“s2”), s3(“s3”) ;
	s0 >> !cnd(start)	>>	s0 ;
	s0 >> cnd(start)	>> sfg0	s1 ;
	s1 >> allways	>> sfg1 >>	s2 ;
	s2 >> allways	>> sfg1 >>	s3 ;
	s3 >> allways	>> sfg2 >>	s0 ;
	}

[0671] As an aid to interpret the C++ code, the equivalent behavior is shown in FIG. 8. The behavior is modeled as a 4-cycle description. Three SFG's (13,14,15) are needed, in addition to a 4-state controller (16). The controller is modeled as a Mealy machine.

[0672] The C++ description also illustrates some of the main contributions of OCAPI: register-transfer level aspects (signals, clocks, registers), as well as dataflow aspects simulation queues) are freely intermixed and used as appropriate. By making use of C++ operator overloading and classes, these different design concepts are represented in a compact syntax format. Compactness is a major design issue.

[0673] Having this specification, we have all information to proceed with the detailed architectural design of the actor. This is however only part of the system design solution: we are also interested in how to incorporate the cycle-true result in the overall system.

Time Verification

[0674] The introduction of time (clock cycles) in the simulation uses an expectation-based approach. It allows to use either a high level or else an SFG-type description of the actor, and simulate the complete system clock-cycle true. The simulation helps the designer in finding whether his ‘high-level’ description matches the SFG description, and secondly, whether the system is realizable.

[0675] A summary of the expectation based simulation is given in FIG. 10 and is used to illustrate the ideas mentioned below.

[0676] This is a different approach then when analysis is used (e.g. the evaluation of a compile-time schedule and token lifetimes) to force restrictions onto the actor implementation. This traditional approach gives the designer no clue on whether he is actually writing down a reasonable description.

[0677] Each token in the simulation is annotated with a time when it is created: the token age. Initial tokens are born at age 0, and grow older as they proceed through the dataflow graph. The unit of time is the clock cycle.

[0678] Additionally, each queue in the simulation holds a queue age (say, ‘the present’) that is used to check the causality of the simulation: a token entering a queue should not be younger than this boundary. A queue is only able to delay tokens (registers), and therefore can only work with tokens that are older than the queue age.

[0679] If such a consistency violation is detected, a warning message is issued and the token age is adapted to that of the queue. Otherwise, the time boundary of the queue is updated with the token age after the token is installed on the queue.

[0680] The queue age is steered by the actor that drives it. For each actor the designer formulates an iteration time. The iteration time corresponds the cycle budget that the designer expects to need for the detailed actor description. Upon each actor firing, the queues driven by the actor are aged with the iteration time.

[0681] At the same time, the actor operations also increase the age of the tokens they process. For normal operations, the resulting token age is equal to the maximum of the operand token ages. For registered signals (only present in SFG-level actor descriptions), the token age is increased by one. Besides aging by operation, aging inside of the queues is also possible by attaching a travel delay to each queue.

[0682] Like the high-level actor description, a queue is also annotated with a number of expectations. These annotations reflect what the implementation of the queue as a set of communication busses should look like.

[0683] A communication bus contains one or more registers to provide intermediate storage, and optionally also a handshake-protocol circuit. A queue then maps to one or more (for parallel communication) of these communication busses.

[0684] The expectations for a simulation queue are:

• • [0685] The token concurrency, that expresses how many tokens of the same age can be present on one queue. To communicate a MATLAB vector annotated with 8@2 for example requires two communication busses. This is reflected in the high level queue model by setting the token concurrency to two.
  • [0686] In case the token concurrency is 1, it can be required that subsequent tokens are separated by a determined number of clock cycles. In combination with the travel delay, this determines how many registers are needed on a communication bus. This expectation is called the token latency.

[0687] Example implementations for different expectations are shown in FIG. 9.

[0688] When the token concurrency is different from one, the token latency cannot be bigger than one. If it would, then the actor that provides the tokens can be designed more effectively using hardware sharing, and thus reducing the token concurrency.

[0689] A summary of the expectation based simulation is put as follows. First, there are several implicit adaptations to token ages and queue ages.

• • [0690] An actor description increases the queue age upon each actor iteration with the iteration time.
  • [0691] A queue increases the age of communicated tokens with the travel delay.
  • [0692] An SFG description increases token ages through the operations. The token age after a register is increased by one, all other operations generate a token with age equal to the maximum of the operand ages.

[0693] The set of operations that modify the token age are referred to as token aging rules.

[0694] Next, a number of checks are active to verify the consistency of the simulation.

• • [0695] A token age cannot be younger (smaller) then a queue age.
  • [0696] The token concurrency on a queue cannot be exceeded.
  • [0697] The token latency on a queue cannot be exceeded.

[0698] A successful clock-cycle true simulation should never fail any of these checks. In the case of such success, the expectations on the queue can be investigated more closely to devise a communication bus for it. In this description we did not mention the use of handshake protocol circuits. A handshake protocol circuit can be used to synchronize tokens of different age at the input of an actor.

Implementation

[0699] The current library of OCAPI allows to describe a system in C++ by building on a set of basic classes.

• • [0700] A simulation queue class that transports a token class and allows to perform expectation-checks.
  • [0701] An SFG/FSM class that allows clock cycle true specification, simulation and code generation.
  • [0702] A token class that allows to simulate both floating point-type representation and fixed point type representation.

[0703] One can simulate the MATLAB data-vector data-type with C++ simulation queues. For the common MATLAB operations, one can develop a library of SFG descriptions that reflect different flavors of parallelism. For instance, a C++ version of the description

[0704]

	% input data
	in = [1 2 1 3 3 4 1 2] ;
	% spreading code
	c = [1 −1 1 −1] ;
	% correlate
	ot = corr (in, c)
	% find correlation peak
	[max, maxpos] = max (ot) ;

looks, after scaling of the parallelism and defining the actor boundaries, like

[0705]

	FB in, ot, maxp ;
	in.delay(1,0) ;	//iteration time, travel delay
	ot.delay(1,0) ;
	maxp.delay(4,0) ;
	in.expect(1,1) ;	//travel time, concurrency,
		latency
	ot.expect(1,1) ;
	maxp.expect(1,4) ;
	in = vector(1, 2, 1, 3, 3, 4, 1, 2) ;
	ot = corr(8, 4, in, vector(1, −1, 1, −1))
	maxp = maxpos(4, ot) ;

[0706] This C++ description contains all information necessary to simulate the system in mind at clock cycle true level and to generate the synthesizable code for the system and the individual actors.

[0707] Thus, the data-flow level has become transparent—it is not explicitly seen by the designer but rather it is implied through the expectations (pragma's) and the library.

[0708] An example of processor design is given next to experience hardware design when using OCAPI.

[0709] The task is to design a 4-tap correlator processor that evaluates a correlation value each two cycles. One coefficient of the correlation pattern needs to be programmable and needs to be read in after a control signal is asserted. The listing in FIG. 11 gives the complete FSMD model of this processor.

[0710] The top of the listing shows how types are declared in OCAPI. For example, the type T_sample is 8 bits wide and has 6 bits beyond the binary point.

[0711] For such a type declaration, a signed, wrap-around and truncating representation is assumed by default. This can be easily changed, as for instance in

[0712]

	// floating point
	dfix T_sample;
	//unsigned
	dfix T_sample(8, 6, ns);
	//unsigned, rounding
	dfix T_sample(8, 6, ns, rd);

[0713] Below the type declarations we see coefficient declarations. These are specified as plain double types, since they will be automatically quantized when read in into the coefficient registers. It is possible to intermix existing C/C++ constructs and types with new ones.

[0714] Following the coefficients, the FSMD definition of the correlator processor is shown. This definition requires: the specification of the instruction set that is processed by this processor, and a specification of the control behavior of the processor. For each of these, OCAPI uses dedicated objects.

[0715] First, the instruction set is defined. Each instruction performs data processing on signals, which must be defined first. The definitions include plain signals (sample_in and corr_out), registers (accu), and register arrays (coef[ ] and sample[ ]).

[0716] Next, each of the instructions are defined. A definition is started by creating a SFG object. All signal expressions that come after such an SFG definition are considered to make up part of it. A SFG definition is closed simply by defining a new SFG object.

[0717] The first instruction, initialize_coefs, initializes the coefficient registers coef[ ]. The for loop allows to express the initialization in a compact way. Thus, the initialize_coefs instruction is also equivalent to

[0718]

	coef[0] = W(T_coef, hardwired_coef[0]);
	coef[1] = W(T_coef, hardwired_coef[1]);
	coef[2] = W(T_coef, hardwired_coef[2]);
	coef[3] = W(T_coef, hardwired_coef[3]);

[0719] The second instruction programs the value of the first coefficient. The new value, coef_in, is read from an input port of the FSMD with the same name. Beyond this port, we are ‘outside’ of the timed FSMD description and use dataflow semantics, and communicate via queues.

[0720] The third and fourth instruction, correl_—1 and correl_—2 describe the two phases of the correlation. It is very easy to express complex expressions just by using C++ operators. Also, a cast operation is included that limits the precision of the intermediate expression result. Although this is for minor importance for simulation, it has strong influence on the hardware synthesis result.

[0721] The instruction read_sample shifts the data delay line. In addition to a for loop, an if expression is used to express the boundary value for the delay line. Use of simple C++ constructs such as these allow to express signal flow graph structure in a compact an elegant way. It is especially useful in parametric design.

[0722] The last instruction, read_control, reads in the control value that will decide whether the first correlation coefficient needs to be refreshed.

[0723] Below all SFG definitions, the control behavior of the correlator processor is described. An FSM with tree states is defined, using one initial state rst, and two normal states phase_—1 and phase_—2. Next, four transitions are defined between those three states. Each transition specifies a start state, the transition condition, a set of instructions to execute, and a target state. For a designer used to finite state machine specification, this is a very compact and efficient notation.

[0724] The transition condition always is always true, while a transition condition like cnd(load) will be true whenever the register load contains a one.

[0725] The resulting fsm description is returned to OCAPI by the last return statement. The simulator and code generator can now process the object hierarchy in order to perform semantical checks, simulation, and code generation.

[0726] The translation to synthesizable VHDL and Cathedral-3 code is automatic and needs no extra designer effort. The resulting circuit for datapath and controller is shown in FIG. 12. The hierarchy of the generated code that is provided by OCAPI is also indicated. Each controller and datapath are interlinked using a link cell. The link cell itself can be embedded into an automatically generated testbench or also in the system link cell that interconnects all components.

[0727] The design of a 75 Kgate DECT transceiver is used as another example (FIG. 13).

[0728] The design consists of a digital radiolink transceiver ASIC, residing in a DECT base station (20) (FIG. 13). The chip processes DECT burst signals, received through a radio frequency front-end RF (21). The signals are equalized (22) to remove the multipath distortions introduced in the radio link. Next, they are passed to a wire-link driver DR (23), that establishes communication with the base station controller BSC (24). The system is also controlled locally by means of a control component CTL (25).

[0729] The specifications that come with the design of the digital transceiver ASIC in this system are as follows:

• • [0730] The equalization involves complex signal processing, and is described and verified inside a high level design environment such as MATLAB.
  • [0731] The interfacing towards the control component CTL and the wire-link driver DR on the other hand is described as a detailed clock-cycle true protocol.
  • [0732] The allowed processing latency is, due to the real time operation requirements, very low: a delay of only 29 DECT symbols (25.2 μseconds) is allowed. The complexity of the equalization algorithm, on the other hand, requires up to 152 data multiplies per DECT symbol to be performed. This implies the use of parallel data processing, and introduces a severe control problem.
  • [0733] The scheduled design time to arrive from the heterogeneous set of specifications to the verified gate level netlist, is 18 person-weeks.

[0734] The most important degree of freedom in this design process is the target architecture, which must be chosen such that the requirements are met. Due to the critical design time, a maximum of control over the design process is required. To achieve this, a programming approach to implementation is used, in which the system is modelled in C++. The object oriented features of this language allows to mix high-level descriptions of undesigned components with detailed clock-cycle true, bit-true descriptions. In addition, appropriate object modelling allows the detailed descriptions to be translated to synthesizable HDL automatically. Finally, verification testbenches can be generated automatically in correspondence with the C++ simulation.

[0735] The result of this design effort is a 75 Kgate chip with a VLIW architecture, including 22 datapaths, each decoding between 2 and 57 instructions, and including 7 RAM cells. The chip has a 194 die area in 0.7 CMOS technology.

[0736] The C++ programming environment allows to obtain results faster then existing approaches. Related to register transfer design environments such as, it will be shown that C++ allows to obtain more compact, and consequently less error prone descriptions of hardware. High level synthesis environments could solve this problem but have to fix the target architecture on beforehand. As will be described in the case of the DECT transceiver design, sudden changes in target architecture can occur due to hard initial requirements, that can be verified only at system implementation.

[0737] First, the system machine model is introduced This model includes two types of description: high-level untimed ones and detailed timed blocks. Using such a model, a simulation mechanism is constructed. It will be shown that the proposed approach outperforms current synthesis environments in code size and simulation speed. Following this, HDL code generation issues and hardware synthesis strategies are described.

System Machine Model

[0738] Due to the high data processing parallelism, the DECT transceiver is best described with a set of concurrent processes. Each process translates to one component in the final system implementation.

[0739] At the system level, processes execute using data flow simulation semantics. That is, a process is described as an iterative behavior, where inputs are read in at the start of an iteration, and outputs are produced at the end. Process execution can start as soon as the required input values are available.

[0740] Inside of each process, two types of description are possible. The first one is a high level description, and can be expressed using procedural C++ constructs. A firing rule is also added to allow dataflow simulation.

[0741] The second flavour of processes is described at register transfer level. These processes operate synchronously to the system clock. One iteration of such a process corresponds to one clock cycle of processing.

[0742] For system simulation, two schedulers are available. A dataflow scheduler is used to simulate a system that contains only untimed blocks. This scheduler repeatedly checks process firing rules, selecting processes for execution as their inputs are available.

[0743] When the system also contains timed blocks, a cycle scheduler is used instead. The cycle scheduler manages to interleave execution of multi-cycle descriptions, but can incorporate untimed blocks as well.

[0744] FIG. 14 shows the front-end processing of the DECT transceiver, and the difference between data-flow and cycle scheduling. At the top, the front-end processing is seen. The received signals are sampled by and A/D, and correlated with a unique header pattern in the header correlator HCOR. The resulting correlations are detected inside a header detector block HDET. A simulation with high level descriptions uses the dataflow scheduler. An example dataflow schedule is seen in the middle of the figure. The A/D high level description produces 3 tokens, which are put onto the interconnect communication queue. Next, the correlator high level description can be fired three times, followed by the detector processing.

[0745] When a cycle true description of the A/D and header correlator on the other hand is available, this system can be simulated with the cycle scheduler as shown on the bottom of the figure. This time, behavior of the A/D block and correlator block are interleaved. As shown for the HCOR block, executions can take multiple cycles to perform. The remaining high level block, the detector, contains a firing rule and is executed as required. Related to the global clock grid, it appears as a combinatorial function.

[0746] Detailed process descriptions reflect the hardware behavior of a component at the same level of the implementation. To gain simulation performance and coding effort, several abstractions are made.

[0747] Finite Wordlength effects are simulated with a C++ fixed point library. It has been shown that the simulation of these effects is easy in C++. Also, the simulation of the quantization rather than the bitvector representation allows significant simulation speedups

[0748] The behavior is modelled with a mixed control/data processing description, under the form of a finite state machine coupled to a datapath. This model is common in the synthesis community. In high throughput telecommunications circuits such as the ones in the DECT transceiver ASIC, it most often occurs that the desired component architecture is known before the hardware description is made. The FSMD model works well for these type of components.

[0749] The two aspects, wordlength modelling and cycle true modelling, are available in the programming environment as separate class hierarchies. Therefore, fixed point modelling can be applied equally well to high level descriptions.

[0750] As an illustration of cycle true modelling, a part of the central VLIW controller description for the DECT transceiver ASIC is shown in FIG. 15. The top shows a Mealy type finite state machine (30). As actions, the signal flowgraph descriptions (31) below it are executed. The two states execute and hold correspond to operational and idle states of the DECT system respectively. The conditions are stored in registers inside the signal flowgraphs. In this case, the condition holdrequest is related to an external pin.

[0751] In execute state, instructions are distributed to the datapaths. Instructions are retrieved out of a lookup table, addressed by a program counter. When holdrequest is asserted, the current instruction is delayed for execution, and the program counter PC is stored in an internal register. During a hold, a nop instruction is distributed to the datapaths to freeze the datapath state. As soon as holdrequest is removed, the stored program counter holdpc addresses the lookup table, and the interrupted instruction is issued to the datapaths for execution.

Signals and Signal Flow Graphs

[0752] Signals are the information carriers used in construction of a timed description. Signals are simulated using C++ sig objects. These are either plain signals or else registered signals. In the latter case the signals have a current value and next value, which is accessed at signal reference and assignment respectively. Registered signals are related to a clock object clk that controls signal update. Both types of signals can be either floating point values or else simulated fixed point values.

[0753] Using operations, signals are assembled to expressions. By using the overloading mechanism as shown in FIG. 16, the parser of the C++ compiler is reused to construct the signal flowgraph data structure.

[0754] An example of this is shown in FIG. 17. The top of the figure shows a C++ fragment (40). Executing this yields the data structure (41) shown below it. It is seen that

• • [0755] the signal flowgraph consists both of user defined nodes and operation nodes. Operation nodes keep track of their operands through pointers. The user defined signals are atomic and have null operand pointers.
  • [0756] The assignment operations use reversed pointers allowing to find the start of the expression tree that defines a signal.

[0757] A set of sig expressions can be assembled in a signal flow graph (SFG). In addition, the desired inputs and outputs of the signal flowgraph have to be indicated. This allows to do semantical checks such as dangling input and dead code detection, which warn the user of code inconsistency.

[0758] An SFG has well defined simulation semantics and represents one clock cycle of behavior.

Finite State Machines

[0759] After all instructions are described as SFG objects, the control behavior of the component has to be described. We use a Mealy-type FSM model to do this.

[0760] Again, the use of C++ objects allow to obtain very compact and efficient descriptions. FIG. 18 shows a graphical and C++-textual description of the same FSM. The correspondence is obvious. To describe an equivalent FSM in an event driven HDL, one usually has to follow the HDL simulator semantics, and for example use multi-process modelling. By using C++ on the other hand, the semantics can be adapted depending on the type of object processed, all within the same piece of source code.

Architectural Freedom

[0761] An important property of the combined control/data model is the architectural freedom it offers. As an example, the final system architecture of the DECT transceiver is shown in FIG. 19. It consists of a central (VLIW) controller (50), a program counter controller (51) and 22 datapath blocks. Each of these are modelled with the combined control/data processing shown above. They exchange data signals that, depending on the particular block, are interpreted as instructions, conditions or signal values. By means of these interconnected FSMD machines, a more complex machine is constructed.

[0762] It is now motivated why this architectural freedom is necessary. For the DECT transceiver, there is a severe latency requirement. originally, a dataflow target architecture was chosen (FIG. 20), which is common for this type of telecommunications signal processing. In such an architecture, the individual components are controlled locally and data driven. For example, the header detector processor signals a DECT header start (a correlation maximum), as soon as it is sure that a global maximum is reached.

[0763] Because of the latency requirement however, extra delay in this component cannot be allowed, and it must signal the first available correlation maximum as a valid DECT header. In case a new and better maximum arrives, the header detector block must then raise an exception to subsequent blocks to indicate that processing should be restarted. Such an exception has global impact. In a data driven architecture however, such global exceptions are very difficult to implement. This is far more easy in a central control architecture, where it will take the form of a jump in the instruction ROM. Because of these difficulties, the target architecture was changed from data driven to central control. The FSMD machine model allowed to reuse the datapath descriptions and only required the control descriptions to be reworked. This architectural change was done during the 18-week design cycle.

The Cycle Scheduler

[0764] Whenever a timed description is to be simulated, a cycle scheduler is used instead of a dataflow scheduler. The cycle scheduler creates the illusion of concurrency between components on a clock cycle basis.

[0765] The operation of the cycle scheduler is best illustrated with an example. In FIG. 21, the simulation of one cycle in a system with three components is shown. The first two, components 1 (60) and 2 (61), are timed descriptions constructed using fsm and sfg objects. Component 3 (62) on the other hand is decribed at high level using a firing rule and a behavior. In the DECT transceiver, such a loop of detailed (timed) and high level (untimed) components occurs for instance in the RAM cells that are attached to the datapaths. In that case, the RAM cells are described at high level while the datapaths are described at clock cycle true level.

[0766] The simulation of one clock cycle is done in three phases. Traditional RT simulation uses only two; the first being an evaluation phase, and the second being a register update phase.

[0767] The three phases used by the cycle scheduler are a token production phase, an evaluation phase and a register update phase.

[0768] The three-phase simulation mechanism is needed to avoid apparent deadlocks that might exist at the system level. Indeed, in the example there is a circular dependency in between components 1, 2, and 3, and a dataflow scheduler can no longer select which of the three components should be executed first. In dataflow simulation, this is solved by introducing initial tokens on the data dependencies. Doing so would however require us to devise a buffer implementation for the system interconnect, and introduce an extra code generator in the system.

[0769] The cycle scheduler avoids this by creating the required initial tokens in the token production phase. Each of the phases operates as follows.

• • [0770] [0] Each the start of clock cycle, the sfg descriptions to be executed in the current clock cycle are selected. In each fsm description, a transition is selected, and the sfg related to this transition are marked for execution.
  • [0771] [1] Token production phase. For each marked sfg, look into the dependency graph, and identify the outputs that solely depend on registered signals and/or constant signals. Evaluate these outputs and put the obtained tokens onto the system interconnect.
  • [0772] [2] (a) Evaluation phase (case a). In the second phase, schedule marked sfg and untimed blocks for execution until all marked sfg have fired. Output tokens are produced if they are directly dependent on input tokens for timed sfg descriptions, or else if they are outputs of untimed blocks.
  • [0773] [2] (b) Evaluation phase (case b). Outputs that are however only dependent on registered signals or constants will not be produced in the evaluation phase.
  • [0774] [3] Register update phase. For all registered signals in marked sfg, copy the next-value to the current-value.

[0775] The evaluation phase of the three-phase simulation is an iterative process. If a pre-set amount of iterations have passed, and there are still unfired components, then the system is declared to be deadlocked. This way, the cycle scheduler identifies combinatorial loops in the system.

Code Generation and Simulation Strategy

[0776] The clock-cycle true, bit-true description of system components serves a dual purpose. First, the descriptions have to be simulated in order to validate them. Next, the descriptions have also to be translated to an equivalent, synthesizable HDL description.

[0777] In view of these requirements, the C++ description itself can be treated in two ways in the programming environment. In case of a compiled code approach, the C++ description is translated to directly executable code. In case of an interpreted approach, the C++ description is preprocessed by the design system and stored as a data structure in memory.

[0778] Both approaches have different advantages and uses. For simulation, execution speed is of primary importance. Therefore, compiled code simulation is needed. On the other hand, HDL code generation requires the C++ description to be available as a data structure that can be processed by a code generator. Therefore, a code generator requires an interpreted approach.

[0779] We solve this dual goal by using a strategy as shown in FIG. 22. The clock-cycle true and bit-true description of the system is compiled and executed. The description uses C++ objects such as signals and finite state machine descriptions which translate themselves to a control/data flow data structure.

[0780] This data structure can next be interpreted by a simulator for quick verification purposes. The same data structure is also processed by a code generator to yield two different descriptions.

[0781] A C++ description can be regenerated to yield an application-specific and optimized compiled code simulator. This simulator is used for extensive verification of the design because of the efficient simulation runtimes. A synthesizable HDL description can also be generated to arrive at a gate-level implementation.

[0782] The simulation performance difference between these three formats (interpreted C++ objects, compiled C++, and HDL) is illustrated in table 1. Simulation results are shown for the DECT header correlator processor, and also the complete DECT transceiver ASIC.

[0783] The C++ modelling gains a factor of 5 in code size (for the interpreted-object approach) over RT-VHDL modeling. This is an important advantage given the short design cycle for the system. Compiled code C++ on the other hand provides faster simulation and smaller process size then RT-VHDL.

[0784] For reference, results of netlist-level VHDL and Verilog simulations are given.

[0785]

TABLE 1
			Source	Simulation	Process
	Size		Code	Speed	Size
Design	(Gates)	Type	(#lines)	(cycles/s)	(Mb)
HCOR	6K	C++(interpreted	230	69	3.8
		obj)
		C++(compiled)	1700	819	2.7
		VHDL (RT)	1600	251	11.9
		VHDL (Netlist)	77000	2.7	81.5
DECT	75K	C++(interpreted	8000	2.9	20
		obj)
		C++(compiled)	26000	60	5.1
		Verilog	59000	18.3	100
		(Netlist)

Synthesis Strategy

[0786] Finally, the synthesis approach that was used for the DECT transceiver is documented. As shown in FIG. 1D, the clock-cycle true, bit-true C++ description can be translated from within the programming environment into equivalent HDL.

[0787] For each component, a controller description and a datapath description is generated, in correspondence with the C++ description. This is done because we rely on separate synthesis tools for both parts, each one optimized towards controller or else datapath synthesis tasks.

[0788] For datapath synthesis, we rely on the Cathedral-3 back-end datapath synthesis tools, that allow to obtain a bitparallel hardware implementation starting from a set of signal flowgraphs. These tools allow operator sharing at word level, and result in run times less than 15 minutes even for the most complex, 57-instruction data path of the DECT transceiver.

[0789] Controller synthesis on the other hand is done by logic synthesis such as Synopsys DC. For pure logic synthesis such as FSM synthesis, this tool produces efficient results. The combined netlists of datapath and controller are also post-optimized by Synopsys DC to perform gate-level netlist optimizations. This divide and conquer strategy towards synthesis allows each tool to be applied at the right place.

[0790] During system simulation, the system stimuli are also translated into testbenches that allow to verify the synthesis result of each component. After interconnecting all synthesized components into the system netlist, the final implementation can also be verified using a generated system testbench.

[0791] A QAM transmission system, that includes a transmitter, a channel model, and a receiver is designed.

System Specification

[0792] A system specification in OCAPI is an executable model: an executable file, that can be run as a software program on a computer. The principle of executable specification, as it is called, is very important for system design. It allows one to check your specification using simulations. In this case, we are designing a QAM transmission system. A full communications system contains a transmitter, a channel model, and a receiver. The ensemble of the transmitter, channel model and receiver organized as an executable specification is also called an end-to-end executable specification. The term end-to-end clearly indicates that the simulation starts with a user message, and ends with a (received) user message. In between, the complete digital transmission is modeled, as shown in FIG. 23.

[0793] In this text, the complete transmission system will be developed. The development of a component in such a system is never a one-shot process. Rather, development proceeds through a design flow: a collection of subsequent design levels connected by ‘natural’ design tasks. For a modem, the typical design levels are:

• • [0794] a statistical level, to do high level explorations of algorithms. In OCAPI, this level is called the link level.
  • [0795] a functional level, to assemble selected algorithms into a single operational modem. In OCAPI, this level is called the algorithm level.
  • [0796] a structural level, to represent the modem as a machine that executes a functional description. In OCAPI, this level is called the architecture level. Each of these levels has an own set of requirements. Statistical requirements can be for example a bit error rate or a cell loss ratio. Functional requirements are for instance the set of modulation schemes to support. Finally, structural requirements are requirements like type of interfaces, or preselected architectures.

[0797] Arranging the requirements besides the design levels yields the design flow, as shown in FIG. 1B. The dashed box contains the levels that will be coded in C++-OCAPI. The upper level (the statistical one) is described in a language like Matlab. It is not included in this text: We will start from a complete functional specification. The functional specification is given herebelow in part A.

Design Flow in OCAPI-C++

[0798] Overall Design Flow

[0799] A design flow with OCAPI looks, from a high level point of view, as shown in FIG. 1C. The initial specification is an architecture model, constructed in C++. Through the use of refinement, we will construct an architecture model out of it. Next, relying on code generation, we obtain a synthesizable architecture model. This model can be converted to a technology-mapped architecture in terms of gates. OCAPI is concerned with the C++ layers of this flow, an in addition takes care of code generation issues.

[0800] Algorithmic Models

[0801] The algorithmic models in OCAPI use the dataflow computational model. The construction of this code by small examples selected out of Part B (below) is discussed.

[0802] First, we consider the construction of an actor. An actor is a subalgorithm out of a dataflow system model. In OCAPI, each actor is defined by one class. As an example of actor definition, we take the diffenc block out of the transmitter. The include file (3.3) defines a class diffenc (line 10) that inherits from a base class. This inheritance defines the class under definition as a dataflow actor. The dataflow actor defines a constructor, a run method and a reset method. The run method (line 25) is the method that is called when the actor should be executed. This method takes along parameters that include the name (name), the I/O ports (_sym 1, _symb2) and other attributes (_qpsk, _diff_mode). The type FB (Flow-Buffer) is the type of a FIFO queue. Looking at the implementation of run (??, line 26), we distinguish a firing rule in lines 29-30. The getsize( ) method of a queue returns the number of elements in that queue. The firing rule expresses that the run( ) method should return whenever there is no data available in the queue. Otherwise, processing continues as described beyond line 32 (This processing is the implementation of the spec as described in Part A.

[0803] A dataflow system is constructed out of such actors. The system code in 5.3 shows how the diffenc actor is instantiated (lines 57-61). Besides actors, the system code also creates interconnect queues (lines 42-48). By giving these as parameters in the constructor of actors, the required communication links are established. Besides the interconnection of actors, the system code also needs to create a scheduler. This scheduler will repeatedly test firing rules in the actors (by calling their run( ) method). The system scheduler that steers the differential encoder is shown on line 77 of 5.3. After this object is created, all dataflow actors that should be under control of it are “shifted into” it. The scheduler object has a method, run( ), that tries firing all of the actors associated with the schedule just once.

[0804] Architecture Models

[0805] An architecture model expresses the behavior of the algorithmic model in terms of operations onto hardware. The kind of hardware features that affect this depend of course on the target architectural style. OCAPI is intended for a bit-parallel, synchronous style. For this kind of style, two kinds of refinements are necessary: First, the data types need to be expressed in terms of fixed point numbers. Second, the execution needs to proceed in terms of clock cycles. The first kind of refinement is called fixed point modeling. The second kind is called cycle true modeling. These two refinements can be done in any order; for a complete architecture model, both are needed. We first give an example on how fixed point numbers are expressed in C++. Consider the ad block of the transmitter (3.2, line 24-27). The purpose of this block is to introduce a quantization effect, such as for instance would be encountered when the signal passes through an analog-digital or digital-analog converter. In this case, the high level algorithmic model is constructed with a fixed point number in order to perform this quantization. On line 32, an object of type dfix (called indfix) is created. This object represents a fixed point value. The constructor uses three parameters. The first, ‘0’, provides an initial value. The following two (W and L) are parameters that represent the wordlength and fractional wordlength respectively. The operation of the ad block is as follows. When there is information in the input queue, the value read is assigned to the fixed point number indfix. At the moment of assignment, quantization happens, whether or not the input value was a floating point value (The FIFO buffers are actually passing along objects of type dfix, so that floating as well as fixed point numbers can be passed from one block to the other). A next example will show how cycle true modeling is done. We consider the derandomizer function of the receiver (6.4). First, looking at the algorithmic model (line 69-102), we see that the block reads two inputs (byte_in and syncro) and writes one output (byte_out). In between, it performs some algorithmic processing (line 89-97). The architecture model is shown in the define( ) function starting at line 116. The first few lines are type definitions and signal declarations. Next, four instructions are defined (line 143-179), and a controller which sequences these instructions is specified (line 184-195). The architecture model makes heavily use of macros to ease the job of writing code. All of these are explained above. The goal of the define( ) function is to define an object hierarchy consisting of signals, expressions, states, etc . . . that represents the cycle true behavior of a processor. At the top of the hierarchy is a finite state machine object. The member function fsm( ) (line 106) returns this object (which is a data member of the derandomizer class). The system integration of the derandomizer (5.3, line 169-176) is the same for the algorithmic and architecture model. The selection between algorithmic and architecture model is done by giving a system scheduler either a base object (as in line 186) or else the fsm object for simulation (as in line 206). Remember that the algorithmic model derives creates a class that derives from the base object; while an architecture model defines a finite state machine object.

[0806] Code Generation

[0807] Finally we indicate the output of the code generation process. When an architecture model is constructed, several code generators can be used. OCAPI currently can generate RT-VHDL code directly, or else also Cathedral-3 dsfg code. When the member function generate( ) of a system scheduler is called, Cathedral-3 code will be produced, along with the required system link cells. The member function vhdlook( ) on the other hand produces RT-VHDL code. In this example, we have used the vhdlook( ) method (5.2, line 401). We consider the derandomizer block in the receiver. The first place where this appears in the generated code is in the system netlist (6.13, line 70 and line 143). Next, we can find the definitions of the block itself: its entity declaration (6.14), the RTL code (6.15), and a mapping cell from the fixed-point VHDL type FX to the more common VHDL type std_logic (6.16). By using this last mapping cell, we can also hook up the VHDL code for derand in a generated testbench (6.17). This testbench driver reads stimuli recorded during the C++ simulation and feeds them into the VHDL simulation.

Part A: System Specification

[0808] System Contents

[0809] The end-to-end model of the QAM transmission system under consideration is shown in FIG. 23. It consists of four main components:

• • [0810] A byte generator GEN (201)
  • [0811] A transmitter TX. (203)
  • [0812] A channel model CHAN. (205)
  • [0813] A receiver RX. (207)

[0814] The byte generator generates a sequence of random bytes. These are modulated inside of the transmitter to a QAM signal. The channel model next introduces distortions in the signal, similar to those occurring in a real channel. Finally, the receiver demodulates the signal, returning a decoded byte sequence. If no bit errors occur, then this sequence should be the same as the one created by the byte generator.

[0815] Next, the detailed operation of the transmitter, the channel and the receiver is discussed. For the internal construction of a component, one might however still refer to FIG. 24.

[0816] Transmitter Specification

[0817] The Transmitter includes

• • [0818] rnd: A randomizer, which transforms a byte sequence into a pseudorandom byte sequence. This is done because of the more regular spectral properties of a rando mized (or ‘whitened’) byte sequence.
  • [0819] tuple: A tuplelizer, which chops the transmitted bytes into QAM/QPSK symbols.
  • [0820] diffenc: A differential encoder which applies differential encoding to the symbols.
  • [0821] map: A QAM symbol mapper, which translates QAM symbols to I/Q pulse sequence s.
  • [0822] shape: A pulse shaper, which transforms the pulse sequences to a continous wave. In digital implementation, the temporal ‘continuity’ is achieved by applying oversampling.
  • [0823] da: Finally, there is a block which applies quantization to the signal. This block simulates the effect of a digital-to-analog converter.

[0824] The transmitter reads in a byte sequence, and randomizes this with a pseudorandom byte sequence. The sequence contains a synchronization word to align the receiver derandomizer to the transmitter randomizer. The pseudorandom sequence is generated by exoring a bitstream with a bitstream produced by a linear feedback shift register (LF SR). The LFSR produces a bitstream according to the polynomial g(x)=1+x⁵+x⁶. It next feeds the bytes to a tuplelizer that generates symbols out of the byte sequence according to the following scheme.

[0825] Given bits b7 b6 b5 b4 b3 b2 b1 b0,

[0826]

Bit position	QAM16	QPSK
b7	I symbol 0	I[1]symbol 0
b6	Q symbol 0	I[0]symbol 0
b5	I symbol 1	Q[1]symbol 0
b4	Q symbol 1	Q[0]symbol 0
b3	I symbol 2	I[1]symbol 1
b2	Q symbol 2	I[0]symbol 1
b1	I symbol 3	Q[1]symbol 1
b0	Q symbol 3	Q[0]symbol 1

[0827] The symbols values are next fed to the differential encoder that generates a diff encoded symbol sequence:

• [0828] i=(((˜(a^b)) & (a^glbIstate))|((a^b) & (b^glbQstate))) &1;
• [0829] q=(((˜(a^b)) & (b^glbQstate))|((a^b) & (a^glbIstate))) &1;
  with i and q the output msbs of the differentially encoded symbol; glbIstate, glbQstate the previous values of i and q; and a and b the inputs msbs of the input symbol. The lsbs are left untouched (only for qam16) The differentially encoded symbol sequence is next mapped to the actual symbol value using the following constellation for QPSK.

[0830]

QVal/Ival	−3	+3
+3	2	0
−3	3	1

[0831] For QAM16, the following constellation will be used

[0832]

QVal/Ival	−3	−1	1	+3
+3	11	9	2	3
+1	10	8	0	1
−1	14	12	4	6
−3	15	13	5	7

[0833] After mapping, the resulting complex sequence is pulse shaped. A RRC shaping filter with oversampling n=4 is taken, with the rolloff factor set at r=0.3. After pulse shaping, the sequence is upconverted to fc=fs/4 in the multiplexer block (included in the shaper)

[0834] Channel Model Specification

[0835] The Channel Model contains

• • [0836] FIR filter with programmable taps. The filter is used to simulate linear distortions such as multipath effects.
  • [0837] Noise injection block. The incoming signal is fed into a 20 tap filter. The second, third, fourth and 21th tap of the filter are programmable. Next a noise signal is added to the sequence. The noise distribution is gaussian;
    X1=sqrt(−2 ln*(U1))*cos(2*pi*U2)
    X2=sqrt(−2 ln*(U1))*sin(2*pi*U2)
  • [0838] U1, U2 are independent and uniform [0,1],
  • [0839] X1 and X2 are independent and N(0,1)

[0840] Receiver Specification

[0841] The Receiver includes

[0842]

	lmsff	A feed forward, T/4 spaced LMS Equalizer.
	demap	A demapper, translating a complex signal
		back to a QAM symbol.
	detuple	A detupler, glueing individual symbols back
		to bytes.
	derand	A derandomizer, translating the pseudonoise
		sequence back to an unrandomized sequence.