SCLive

Forewords

Inheritance and multiple inheritances are a pivotal part of interfaces and channels building. Channels in SystemC can be viewed as, a sophisticated form of communication mechanism; Akin to signals or wires in found in other HDLs. However what differentiates SystemC from common HDLs is the ability given to the designer to handcraft his own channels. Furthermore, the use of interfaces enables the separation of the modules’ implementation from the communication scheme used in between modules; This is an essential part of system level modeling.
In this section we will examine the existing predefined SystemC channels as well as look at the creation of user defined primitive and hierarchical channels.

Inheritance and SystemC channels/ Interface loose coupling

For the creation of channels, SystemC borrows from C++ a convenient principle called: “the class interface”. This principle is also known as abstract base classes and is used to define a common interface for related derived classes.
Sometimes in C++, designers want to create a class that will define a set of access methods that should be used by all the classes that will be derived from it; however this class in itself will never be used to create objects. This type of classes is an abstract base class. In SystemC, abstract base classes are used to define all the access methods that a channel should have. Consequently, a channel will be a C++ class derived from an abstract base class and will implement the access methods defined inside its abstract base parent class.

An abstract base class is created in C++ through the definition of one or more pure virtual methods as part of that class. A pure virtual method by definition is a method that will only be implemented inside a derived class from the abstract base class. The semantics for the declaration of a C++ pure virtual method is as follows:

virtual <return_type> function_name (args>)=0;

For instance the following line of code declares a method called bustWrite that takes three input arguments and returns an integer type:

virtual void burstWrite( int destAddress, int numBytes, sc_lv<8> *data ) = 0;

The keyword ‘virtual’ and the ‘=0’ are the essential parts of this declaration as they indicate to the compiler that ‘burstWrite’ is a pure virtual method and therefore, that the class containing this declaration can never be made into an object.

Once one or a number of abstract base classes (interfaces) have been designed, channels can be implemented by simply inheriting one or more of the base classes and implementing their virtual methods.

Following the creation of channel, instances of that channel can be made as any other C++ object.

Primitive Channels

The SystemC language extension distinguishes between two kinds of channel’s implementation: Primitive channels and non-primitive ones also referred as ‘hierarchical’ channels. Primitive channel as their name imply, have restrictions in regards to their construction. Firstly, any primitive channel is derived from both, its associated interfaces and the predefined SystemC class: sc_prim_channel. In addition, a primitive channel is not allowed to contain SystemC structures for instance threads, methods, other channels et caetera. However, although primitive channel appear limited, they uniquely support the SystemC defined ‘request_update()’ and ‘update()’ access methods.

The SystemC library defines a number of versatile primitive channels namely: sc_buffer<T>, sc_fifo<T>, sc_mutex, sc_semaphore, sc_signal<T>, sc_signal_resolved, sc_signal_rv<W>. Those channels can be used individually or can be combined as part of a ‘non-primitive’ channel to create more complex communication mechanisms. These channels can be declared as follows:

// A 32×16 bits elements fifo
sc_fifo<sc_lv<16> > inputFifo(32);
// A 16 bits bus
sc_signal<sc_lv<16> > dataBus;
// A 4 tokens semaphore
sc_semaphore gateKeeper(4); 

The predefined SystemC primitive channels rely on a number of access methods to allow users to examine or alter their internal state. As previously highlighted, those methods are defined inside the channel’s interface. For channels used to hold values, the read() and write() methods are provided. Most predefined channels also support the common methods: print(), dump(), trace()and kind(). The print(), dump() and trace() methods are used to display the internal values of a given primitive channel either using the output stream or a trace file in the case of the trace() method. The purpose of the kind() method is used to identify the current primitive channel by providing it’s kind as a string information.

Along with access and display methods, most primitive channels support queries of events. These methods may be used for two purposes: examining if a specific channel has currently had an event as part of an expression; such as an ‘if’ statement or supply information related to the activity/events of a channel as part of a sensitivity list of a SystemC process. The following examples illustrate the different method available for specific channels.

// A non blocking write on an sc_signal
dataBus.write(”1010111101010101″);
// A blocking write on a variable of type sc_lv
sc_lv<16> busVar = dataBus.read();
while ( inputFifo.num_free() != 0 )
// A blocking write on an sc_fifo
inputFifo.write(dataBus.read());
// A blocking access to an sc_semaphore
gateKeeper.wait();
// A non blocking access to an sc_semaphore
if ( gateKeeper.trywait() != 0 )
cout << “Got a lock !” << endl;

Although SystemC provides us with numerous pre-defined channels, it is also possible to create user defined channels. As mentioned previously, the creation of a channel is achieved through multiple-inheritance. As a rule, a channel inherits from one or more interfaces as well as from either the sc_prim_channel or the sc_channel SystemC class. Sophisticated relationship between channels can be created by multiple inheritances from various interfaces. The following example describes the creation of a simplified version of the SystemC’s semaphore primitive channel.

class simple_semaphore_if : virtual public sc_interface {  // Line 1
public:
   // the classical operations: wait(), trywait(),
   // and post(), lock (take) the semaphore,
   // block if not available
   virtual int wait() = 0; // Line 7

   // lock (take) the semaphore, return -1
   // if not available
   virtual int trywait() = 0; // Line 10

   // unlock (give) the semaphore
   virtual int post() = 0; // Line 13

   // get the value of the semphore
   virtual int get_value() const = 0; // Line 16

   protected:
   // constructor
   simple_semaphore_if() { } // Line 21
};

// Creation of the semaphore Channel
class simple_semaphore: public simple_semaphore_if, public sc_prim_channel
{
   public:
   // constructors
   explicit simple_semaphore( int init_value_ );

   // interface methods
   // lock (take) the semaphore, block if
   // not available
   virtual int wait(); // Line 37

   // lock (take) the semaphore, return -1 if
   // not available
   virtual int trywait(); // Line 40

   // unlock (give) the semaphore
   virtual int post(); // Line 43

   // get the value of the semaphore
   virtual int get_value() const // Line 46
   { return m_value; }

   protected:
   // support methods
   bool in_use() const // Line 53
   { return ( m_value <= 0 ); }

   protected:
   int m_value; // Line 58
   sc_event m_free; // Line 58
};

// constructors
simple_semaphore::simple_semaphore( int init_value_ ) :
   sc_prim_channel( sc_gen_unique_name( “semaphore” ) ),
   m_value( init_value_ ) { }

// interface methods
// lock (take) the semaphore, block if
// not available
int simple_semaphore::wait() {
   while( in_use() )  sc_prim_channel::wait( m_free );
   -- m_value;
   return 0;
}

// lock (take) the semaphore, return -1 if
//not available
int simple_semaphore::trywait() {
   if( in_use() ) {
      return -1;
   }
   -- m_value;
   return 0;
}

// unlock (give) the semaphore
int simple_semaphore::post() {
   ++ m_value;
   m_free.notify();
   return 0;
}

The first part of this example defines the interface for the semaphore’s channel. line 1 declares the interface class named ‘simple_semaphore_if’ and inherits virtually, the sc_interface SystemC predefined class. The virtual inheritance is used here to prevent any problems related to repeated inheritance later-on during the creation of the simple_semaphore channel. The lines 7 to 16 declare pure virtual methods that will be used to access the semaphore channel. Finally the line 21 defines a default constructor for the interface. Following the interface declaration, we find the declaration of the semaphore channel called ‘simple_semaphore’. The class is created through the inheritance of both the simple_semaphore_if interface class and the SystemC sc_prim_channel pre-defined class. The lines 37 to 46 make the previously pure virtual methods defined inside the interface into effective methods inside the channel. A local method is created in line 53 to used to keep track of the current value of the semaphore. Lastly in lines 58 and 59 we find the definitions of internal member variables used by the semaphore channel. The ultimate part of this example implements each individual methods defined inside the semaphore channel.

Hierarchical Channel

Up until now we have considered the simplest form of channels: the primary channels. Although primary channels are of great use, one of the main advantages that SystemC has over conventional HDLs, is its ability to create complex communication mechanisms in the form of hierarchical channels.
Hierarchical channels are used as a convenient way to abstract exchanges between communicating objects such as sc_modules. Commonly hierarchical channels are used as transactors, converting high level commands to RTL style signals and vice versa.

Herarchical channel can be considered in a lot of ways as, more sohisticated forms of sc_modules. As such, a hierarchical channel can contain ports, as well as, a hierarchy of sc_modules or other channels. Furthermore, hierarchical channels may contain SystemC processes such as: SC_METHODS or SC_THREAD. However, unlike their sc_modules counterparts, hierarchical channels will also provide methods for implementing the functions defined inside their associated interfaces.

As we mention earlier, a channel is always created from one or multiple existing interfaces. An interface in SystemC is an abstract class derived from the existing SystemC class: sc_interface.

By definition interfaces will only define pure virtual access methods. These methods will eventually have to be implemented in any deriving channels. The following illustrates the creation of a user defined interface called dma_interface with a methods called: burstWrite() and bustRead().

class dma_interface: virtual public sc_interface {
public:
   virtual void burstWrite(int destAddress, int numBytes, sc_lv<8> *data ) = 0;
   virtual void burstRead(int sourceAddress, int numBytes, sc_lv<8>* data) = 0;
};

Commonly the inheritance of the pre-defined SystemC class: sc_interface, is done as, public and virtual. The virtual mechanism is used as a safety net, preventing the rise of issues related to repeated inheritance when a channel inherits more that one interface.

After the creation of an interface, numerous channels can be created to provide the required implementation of the interfaces’ methods. The creation of a hierarchical channel differs from a primitive one, only by the type of the inherited pre-defined SystemC class: sc_channel. Along with the sc_channel class, one to any number of interface classes can be inherited. The following code illustrates the creation of a channel called: dma_channel, publicly inheriting form the interface class: dma_interface.

class dma_channel: public dma_interface, public sc_channel {
public:
   sc_out_rv<16> address_p;
   sc_inout_rv<8> data_p;
   sc_out_resolved rw_p;
   dma_channel(sc_module_name nm): sc_channel(nm)
      ,address_p(”address_p”)
      ,data_p(”data_p”)
      ,rw_p(”rw_p”)
   { }
   virtual void burstWrite( int destAddress, int numBytes, sc_lv<8> *data );
   virtual void burstRead(int sourceAddress, int numBytes, sc_lv<8>* data);
};

void dma_channel::burstWrite( int destAddress, int numBytes, sc_lv<8> *data ) {
   sc_lv<8> *ite = data;
   for (int i=0; i<numBytes; i++) {
      address_p->write(destAddress++);
      data_p->write( *(ite++) );
      wait(10, SC_NS);
      cout << “Write out ” << data_p->read() << endl;
      rw_p->write(SC_LOGIC_0); // Write pulse
      wait(50, SC_NS);
      rw_p->write(SC_LOGIC_Z);
      address_p->write(”ZZZZZZZZZZZZZZZZ”);
      data_p->write(”ZZZZZZZZ”);
      wait(10, SC_NS);
   }
}

void dma_channel::burstRead(int sourceAddress, int numBytes, sc_lv<8>* data) {
   for (int i=0; i<numBytes; i++) {
      address_p->write(sourceAddress++);
      wait(10, SC_NS);
      rw_p->write(SC_LOGIC_1); // Read pulse
      wait(10, SC_NS);
      *(data++) = data_p->read();
      cout << “Data read ” << data_p->read() << endl;
      wait(40, SC_NS);
      rw_p->write(SC_LOGIC_Z);
      address_p->write(”ZZZZZZZZZZZZZZZZ”);
      data_p->write(”ZZZZZZZZ”);
      wait(10, SC_NS);
   }
}

The first part of this code illustrates the creation of the hierarchical channel called dma_channel. This is done through the multiple inheritance of the sc_channel class and the dma_interface class.
Ports are created to illustrate the flexibility of channels. Those port are initialised inside the constructor of that class.
The last part of the dma_channel class declaration restates the existance of the burstWrite() and burstRead() methods found in the parent class: dma_interface.

Lastly the code illustrates the implementation of the two methods burstWrite() and burstRead(). The code itself is of little importance but it demonstrates how high level transactions can be refined into low-level RTL signal activities.

For the pupose of completeness of this example the following code illustrates how can this channel could be used in a verification environment.
This code shows the creation of a simple testbench sending a burstWrite() and a burstRead() request to a slave RTL memory via the dma_channel.

class test_bench: public sc_module {
   public:
   sc_port<dma_interface> master_port;

   void stimuli()
   {
      sc_lv<8> data_sent[10] = {20, 21, 22, 23, 24, 25,26,27,28,29};
      sc_lv<8> data_rcv[10] = {0,0,0,0,0,0,0,0,0,0};
      master_port->burstWrite(100, 10, data_sent);
      wait(100, SC_NS);
      master_port->burstRead(100, 10, data_rcv);
      for (int i=0; i<10; i++) {
         if (data_sent[i] != data_rcv[i]) {
            cout << data_sent[i] << ” ” << data_rcv[i] << endl;
            cout << “data missmatch” << endl;
         }
      }

   SC_HAS_PROCESS(test_bench);

   test_bench(sc_module_name nm): sc_module(nm) {
      SC_THREAD(stimuli);
   }
};

class rtl_memory: public sc_module {
public:
   sc_in_rv<16> address_p;
   sc_inout_rv<8> data_p;
   sc_in_resolved rw_p;
   sc_lv<8> *mem_arr;
   void run() // sensitive rw_p
   {
      while(true) {
         // read cycle
         if (rw_p->read() == SC_LOGIC_1) {
            data_p->write( *( mem_arr+(sc_uint<16>(address_p->read())) ) );
            // write cycle
         } else if (rw_p->read() == SC_LOGIC_0) {
            *(mem_arr + (sc_uint<16>(address_p->read()))) = data_p->read();
         }
         wait();
      }
   }

   SC_HAS_PROCESS(rtl_memory);
   rtl_memory(sc_module_name nm, int mem_size = 100): sc_module(nm) {
      mem_arr = new sc_lv<8>[mem_size];
      for (int i=0; i< mem_size; i++) {
         mem_arr[i] = sc_lv<8>(0);
      }
   SC_THREAD(run);
      sensitive << rw_p;
   }

   ~rtl_memory() { delete []mem_arr; } };

// Main program
int sc_main(int argc, char* argv[])
{
   sc_set_time_resolution(1, SC_NS);

   sc_signal_rv<16> address_s;
   sc_signal_rv<8> data_s;
   sc_signal_resolved rw_s;

   test_bench tb(”tb”);
   dma_channel transactor(”transactor”);
   rtl_memory uut(”uut”, 1000);

   tb.master_port(transactor);
   transactor.data_p(data_s);
   transactor.rw_p(rw_s);
   transactor.address_p(address_s);
   uut.address_p(address_s);
   uut.data_p(data_s);
   uut.rw_p(rw_s);

   sc_start();
   return 0;
}

Summary

In this section we covered the existance of two kind of channels in SystemC: primitive, hierarchical. The main differences between those two kinds being: primitive channels cannot contain SystemC structural objects (ports, channels, processes) but can use the update() method to implement non blocking update mechanisims; The hierachical channels however can have structural objects but cannot use the update() method.
I addition we illustrated the creation of both a primitive and hierarchical channel as well as their use in the context of a verification environment.

The code for this tutorial can be found HERE.

Et voila !

David Cabanis

Forewords

This tutorial is intended to be a basic introduction to the notions of concurrent processes in SystemC. We will touch on the SystemC ability to dynamically spawn processes.

Processes in SystemC

One of the essential element of the SystemC language is concurrency. This is achieved through what we will call processes. Processes are similar to processes in VHDL or always/initial in Verilog. In principle processes are block of sequential code similar to functions. However, unlike functions, processes are not explicitly called by the user. Processes can in effect be seen as always active; For that reason they are considered to be concurrent.
In the SystemC language there are two distinct kinds of processes namely: SC_METHOD and SC_THREAD. These two kind of processes are typically created statically; In other words they are created prior to the execution of the simulation. Alternatively SystemC allows processes to be created dynamically via a dedicated function called: sc_spawn() .

The SC_THREAD process

As a first step we will look at the SystemC SC_THREAD. A thread by definition is a process that will automatically execute itself at the very start of a simulation run and then will suspend itself for the rest of the simulation.
Threads are versatile processes in the sense that they can be halted at any moment and any number of times during their execution. Interestingly, although the thread is mean to be executed only once throughout the simulation, most of times we want to be able to use threads for the whole duration of the simulation. In order to achieve this, we use a simple ‘trick’ that consist in creating an infinite loop within a thread. Consequently we prevent the thread from ever reaching its sequential end. The following example illustrates how to create a simple thread used to generate a clock generator.

void run() {
   while(true) {
      clock_signal.write(SC_LOGIC_0);
      wait(10, SC_NS);
      clock_signal.write(SC_LOGIC_1);
      wait(10, SC_NS);
   }
}

Three main observations can be made:

• The funtion ‘run’ does not have any input or return parameters. This typical of a SystemC static process such as SC_THREAD or SC_METHOD.
• An infinite while loop is used to prevent the function ‘run’ from ever reaching its end. This does not have to be used for all SC_THREAD implementations; however it is a common occurrence.
• We are using ‘wait’ function calls to force a temporary suspension of the function ‘run’ and the update of the ‘clock_signal’ value. This can only be used by SC_THREAD processes.

Up until now we have only created a C++ member function that will be used as a thread. At this time we will need to indicate to the SystemC simulation kernel that this specific member function (in our case ‘run’) should be treated as a concurrent thread and not as a basic C++ member function. This will be done through an operation called processes registration. To put it simply this consists in enumerating how individual functions should be treated. This simple operation is done inside the constructor of a module. The following code will illustrate how this is achieved.

sample_module(sc_module_name nm):sc_module(nm) {
   SC_THREAD(run);
}

In this example we assume the existence of a SystemC called ’sample_module’. The process registration of the member function ‘run’ as an SC_THREAD is done by passing the name of the selected function (run) as an argument of the macro SC_THREAD.

The SC_METHOD process

The SC_METHOD is in many ways similar to the SC_THREAD. However there are two main differences: SC_METHOD will run more than once by design and methods cannot be suspended by a wait statement during their execution. The following code illustrates this.

void log() {
   count_signal.write(count_signal.read()+1);
   if (count_signal.read() >= 10) {
      cout << "Reached max count" << endl;
      count_signal.write(0);
   }
}

From this sample we can observe that no wait statements are being used. Additionally there is no need for an infinite ‘while’ loop since this is the nature of an SC_METHOD to go back to it’s beginning when it finishes executing.
However not visible inside the code lies a hidden wait statement. That implicit statement is always located on the very last line of the member function and causes the kernel to suspend the SC_METHOD so that its sensitivity list can be evaluated.

Processes Sensitivity

SystemC support two kind of sensitivity: Dynamic and static. For the purpose of this tutorial we will concern ourselves only with that later. Both SC_THREAD and SC_METHOD may have a sensitivity list although they are not obliged to. In the example of the SC_THREAD covered previously we did not use a static sensitivity list; For the SC_METHOD example we will have to.
As seen before the SystemC requires a kernel registration to identify which member function it needs to treat as a SC_METHOD (no explicit suspension allowed) or an SC_THREAD (suspensions allowed). In the case of the method processes we will use the macro SC_METHOD. The following code illustrates the registration of the ‘log’ function as a SC_METHOD.

sample_module(sc_module_name nm):sc_module(nm) {
   SC_METHOD(log);
      sensitive << clock_signal.posedge_event();
}

As can be observed we now have an additional statement indicating that the macro directly above it is sensitive to positive edges of the signal ‘clock_signal’. That sensitivity list will typically contain event objects as well as ports or channels (signals).
Consequently as soon as a positive edge is detected on the signal ‘clock_signal’, the function ‘log’ will be re-executed.

The last point of detail remaining is to indicate to the kernel that this module that we have created will contain concurrent processes. This is done using the macro SC_HAS_PROCESS(moduleName).

The full code for this section on threads, methods and sensitivity list is shown here:

class sample_module: public sc_module
{
	public:
	sc_signal clock_signal;
	sc_signal count_signal;
	SC_HAS_PROCESS(sample_module);
	sample_module(sc_module_name nm):sc_module(nm) {
		SC_THREAD(run);
		SC_METHOD(log);
			sensitive << clock_signal.posedge_event();
	}

	void log() {
		count_signal.write(count_signal.read()+1);
		if (count_signal.read() >= 10) {
			cout << "Reached max count" << endl;
			count_signal.write(0);
		}
	}

	void run() {
		while(true) {
			clock_signal.write(SC_LOGIC_0);
			wait(10, SC_NS);
			clock_signal.write(SC_LOGIC_1);
			wait(10, SC_NS);
		}
	}
};

Dynamic processes

In the first part of this tutorial we have looked at the most common usage of processes in SystemC. Nevertheless there is yet a more sophisticated way of creating processes by spawning functions at will. This method is by far the most flexible and most complete.
The sc_spawn function can be seen as a super-set of the existing SC_THREAD, SC_METHOD macros. It can be used as a straight replacement although the semantics is a bit more involving.
A very basic example of the function sc_spawn is as follows. Assuming that we have an existing member function as such:

void spawned_th(bool value) {
   cout << "spawned_th successfuly spawned with value: "
          << value << endl;
}

A call to spawn that function from within an SC_THREAD would be as such:

sc_spawn( sc_bind(&sample_module::spawned_th, this, true) );

Without going in too much details, this call relies on an additional function sc_bind that is required to attach an existing function. This is done by passing the address (&) of the required member function as an argument. In addition, we need to indicate that this specific function is not a global function (this is also allowed) but a member function of our module hence the use of ‘this’. Finally the function ’spawned_th’ has one input paramenter ‘value’ of type bool therefore we set an actual value (true) inside the ’sc_bind’ call.

A more complete example is as follows:

class sample_module: public sc_module
{
   public:
   sc_signal clock_signal;
   SC_HAS_PROCESS(sample_module);
   sample_module(sc_module_name nm):sc_module(nm) {
      SC_THREAD(run);
   }

   void spawned_th(bool value) {
      cout << "spawned_th successfuly spawned with value: "
             << value << endl;
   }

   void run() {
      sc_spawn( sc_bind(&sample_module::spawned_th, this, true) );

      while(true) {
         clock_signal.write(SC_LOGIC_0);
         wait(10, SC_NS);
         clock_signal.write(SC_LOGIC_1);
         wait(10, SC_NS);
      }
   }
};

As stated earlier this example has been kept very simple for the purpose of clarity. Additional features would be the ability to:

• Specify a return argument for the spawned function
• Specify if the argument passed to the function are reference (as opposed to copies) or constant references
• Spawn a global function instead of a member function
• Being able to specify if the spawned function should be treated as a method or thread(default)
• Being able synchronise thread activities by checking their status (idle or running)

Most of those options can fairly easily be achieved however they fall outside of the scope of this tutorial. Nevertheless to give you a more complete picture of the ’sc_spawn’ function, we will illustrate the spawning of a global function with return arguments and reference input arguments.

#define SC_INCLUDE_DYNAMIC_PROCESSES

int global_th(const bool& in_value, int& out_value) {
   cout << "global_th successfuly spawned with value: "
          << in_value << ", " << out_value << endl;

   if (in_value) {
      out_value++;
      wait(2, SC_NS);
      return 0;
   }
   wait(5, SC_NS);
   return 1;
}

class sample_module: public sc_module
{
   public:
   sc_signal clock_signal;
   SC_HAS_PROCESS(sample_module);
   sample_module(sc_module_name nm):sc_module(nm) {
      SC_THREAD(run);
   }

   void run() {
      int actual_return_value, actual_int_arg = 9;
      bool actual_bool_arg = true;

      sc_spawn( &actual_return_value, sc_bind(&global_th,
                       sc_cref(actual_bool_arg),
                       sc_ref(actual_int_arg) ) );

      while(true) {
         clock_signal.write(SC_LOGIC_0);
         wait(10, SC_NS);
         clock_signal.write(SC_LOGIC_1);
         wait(10, SC_NS);
      }
   }
};

This time the ’sc_spawn’ function has an additional argument for the returned value (passed by pointer) in addition the ‘this’ argument was removed since we are now dealing with a global function. The ’sc_cref’ and ’sc_ref’ functions were used to indicate that we were dealing with constant reference and reference formal arguments.

Conclusions

This tutorial covered the basic uses of SystemC processes. we announced the rules regarding the suspension of processes and looked at the sensitivity list and kernel registrations. Lastly we covered the use of the sc_spawn function and saw how can generic functions can be parallelised at will.

Et voila!

David Cabanis.