SystemC: Key modeling concepts besides TLM to boost your simulation performance

Martin Schnieringer, VaST Systems Technology, Munich Germany
Kevin Brand, VaST Systems Technology, Sydney, Australia

Abstract:

Virtualization of SoC, ECUs and other electronic systems is used to explore (micro-)architectures at system level as well as to develop and verify software early in the design cycle. Key to those Virtual System Prototypes (VSP) is the simulation speed and accuracy dependent on the use case. The latest SystemC standards, such as OSCI TLM 2.0, OCP IP TL2 etc. focus on the standardization of the bus interfaces of peripheral devices connected in a virtual system simulation. Recent proposals have extended these standards with regards to the runtime configuration of the HW simulation models. Nevertheless these standards on different levels of abstraction such as programmer's view (PV), programmer's view with timing (PV-T) etc. do not ensure fast simulation speeds for processor centric systems per se. Those standards do not recommend efficient modeling techniques for the behavior of the simulation IP block. This is important as inefficient modeling of the behavior can easily bring down your simulation speed into the KIPS range (kilo instructions per second) instead of the MIPS range (million instructions per second) requested by the software engineers developing their applications on such a virtual platform. This paper will demonstrate a methodology for modeling your virtual hardware efficiently by minimizing events but still being accurate on cycle boundaries. As an example we use a simple watchdog timer to illustrate the basic concepts of such a SC_METHOD based modeling approach which allows us to maintain simulation speeds in the MIPS range. In the end we will also show that complex systems with a decent number of complex devices can still maintain the promised simulation speeds in the MIPS range.

1. Motivation

Virtual System Prototypes are mainly used by two groups of engineering communities, system architects and software developers. While speed proves to be more important for SW developers, accuracy seems to be more important for system architects. Yet, the SW developer does often also require high degrees of accuracy, for example for real-time critical reactive systems, or multimedia applications with stringent throughput requirements. And also the system architect is moving more and more towards SW-driven architecture analysis and optimization strategies, in which real SW loads (OS, Multimedia, …) are used for stimuli. The typical opinion is, that for both communities, two different sets of models are required, each differentiating by either speed of the model or accuracy. The industry has been moving though strongly towards providing speed and accuracy in integrated solutions, with only the best being able to provide speed and accuracy in one model, so called 1-source models, Still, the question remains at hand, how Virtual System Prototypes can be completed by the individual modeling teams to remain fast and accurate. This paper discusses, how peripheral models of IP blocks should be modeled, so that they enable full Virtual System Prototypes to execute embedded target software at two digit MIPS speeds, yet simulate cycle-accurate at a pin and bus boundary level.

2. Minimizing the Number of Events

The simulation speed depends on many factors such as

• Size of the system to model
• Level of detail of the system model
• Performance of the host machine that is executing the simulation

What counts with regards to the simulation speed for your simulator, is the number of events that have to be processed by your simulation kernel for a given simulation time. So in order to maximize simulation performance, it is all about minimizing the number of events. Less events lead to higher MIPS rates hence shorter development time for the application software developer. It is also important to note that less events need not lead to less granularity for the system architect. There are many different types of events and in the following we discuss these and how to minimize their occurrence whilst maintaining cycle accurate boundaries. The events discussed here include:

• Protocol and Interface related events
• Behavioral events
• Clock events

Protocol and Interface Events

When a real device communicates with other real devices in a system, these communication transitions translate to simulation events in the virtual world. Using transaction level interfaces to represent these is the obvious mean to remove unnecessary events. Function calls – also called Interface Method Calls (IMCs) - are used to communicate over a virtual bus infrastructure. The events we could save and hence do not need to model are the ones typically seen in a pin-level accurate interface such as handshaking signals, read and write strobes etc. Of course minimizing the number of events is possible for all kind of protocols. Consider a serial interface and the number of events caused by shifting a 32 bit data word into the receiver resulting in each bit shift causing an event at the serial clock edge. The number of events could be significantly reduced by transmitting the serial data at once – so to say in parallel - and keep track of the start time of the transfer and end time. From a software developers perspective we are just interested in how long it takes to receive that serial data word and visibility into the partly shifted data is not required. Using this modeling style we reduced the number of events – namely to two, one for the start and one for the end time – but kept the accuracy for application software developer and system architect.
Minimizing the number of interface related events is a good starting point but does not ensure great simulation speeds per se. When a TLM interface is used but the behavior of your device is modeled in a RTL fashion (all internal states are updated each clock cycle) it is obvious that modeling edge clocked states that are not visible from outside results in simulation performance loss. These internal state behavior events are required to be abstracted.

Behavioral Events

To better illustrate the efficient modeling techniques a simple watchdog timer is used. The bus slave device has two memory mapped registers. One to control the watchdog timer and a single count register. The boolean output port should go high for one clock period when the timer register counted down to zero. The inefficient RTL style approach would be to decrement the counter register each clock period and even worse doing that in a clocked SC_THREAD causing a context switch each time.
The most important event for the simulation in terms of the watchdog timer is the boolean output port change. Therefore the port change is only interested in one event at a precise time and that is…. The counter reaching zero at the clock edge that caused it.
To model the behavior of driving the output port at the precise clock edge, a SC_METHOD can be implemented that is sensitive to an sc_event. That event is notified in the simulation future using the notify() method for SystemC events. Effectively we are modeling a counter match event that will happen at a known time. It is quite simple to calculate the time in the future and hence when to trigger the SC_METHOD. The time in the future is:

[watchdog counter value] x [clock period]

The clock period could be communicated by using an input port of type unsigned int or sc_time in the peripheral device. For the later one the sc_trace() must be overloaded to support ports of type sc_in<sc_time>. That approach has the benefit that the time unit of the clock period could be passed as well to avoid confusion when IP blocks of different sources are mixed. In the following example we use an input port of type unsigned int and the time unit is 1 picosecond as that is the default resolution of the SystemC simulator in use.

sc_event MatchEvent;
SC_METHOD(MatchCallback);
sensitive << MatchEvent;
dont_initialize();
MatchEvent.notify(TimerClkPeriod * Value,
SC_PS);
..
void watch_dog::MatchCallback() {
TimerInterrupt.write(0x1);
}

To drive the output port low again we could schedule another SC_METHOD one clock period of time in the future using the same approach. And subsequently another match event could be scheduled in the future for the next count down of the timer. This SC_METHOD based approach is sometimes also called a “Callback” based approach. So far, that was pretty easy to accurately model this interrupt event efficiently, however there are other external actions that have an impact on the behavior of the watchdog model.

The following actions all impact the design:

• Reset of the device
• Access of the memory mapped register by the software target code
• Change of the clock period

The reset of the device is the simplest scenario and pending sc_events are canceled and registers are loaded with their reset values.

sc_in<bool> Reset;

SC_METHOD(ResetChanged);
  sensitive_pos << Reset;
  dont_initialize();

void watch_dog::ResetChanged() {
  //Reset member vars, registers, ports
  TimerControlReg = 0x00;
  //Cancel scheduled events
  MatchEvent.cancel();
}

Depending on the design of model, other actions may need to be taken on this reset event e.g. rereading the clock period of a clock input port. The next action that the model may experience from the software developer’s perspective, is reading and writing to the watchdog’s memory mapped registers especially after the watchdog timer has been enabled. Since this action can happen at any time, the correct count value must be returned to the target software. So the challenge is to reflect the accurate state of the counter even though the current state is not known as we have not modeled the decrementing counter. In order to return the correct counter value a new member variable is introduced that stores the simulation time (sc_time) of the last “update” to the model. The variable is updated whenever an access to a memory mapped register occurs that has an impact on the counter state e.g. a write access to the control register or a read/write access to the counter register. The member variable is named mLastUpdate. At any point in time during the simulation, the correct return value of the counter register can be calculated using the following equation:

value = (NowTime – mLastUpdate) / clock_period

NowTime is the simulation time when the read access to the counter register occurs. To ease modeling calculations later on clock ticks can be used instead of absolute times.

A write access to the watchdog control register is still simple. Per se we cancel all pending events and only re-schedule the “Match” event when the enable bit 0x80 is set. The code fragment below shows a possible implementation of the address decode function:

watch_dog::write(int *data, unsigned int address) {
  double NowTime = sc_simulation_time();
  MatchEvent.cancel();
  //Calculate address offset
  int Offset = (address - m_start_address)
    & ~(sizeof(unsigned int) - 1);
  switch (Offset) {
    case WatchDogCtrlOfst:
      WatchDogCtrl = *data;
      if(WatchDogCtrl & 0x80) { //Schedule
        // an MatchEvent
      MatchEvent.notify((mWatchDogClkPeriod
        * WatchDogValue), SC_PS);
      . ..
      }
. ..

A write access to the watchdog counter register needs the Match event to be re-scheduled.

switch (Offset) {
   . ..
  case WatchDogValueOfst:
    WatchDogValue = *data;
    if(WatchDogCtrl & 0x80) { //Schedule
       //MatchEvent based on new value
    MatchEvent.notify((mWatchDogClkPeriod
* WatchDogValue), SC_PS);
    }

The example above is pretty simple and the bus interface and the watchdog timer rely on the same clock period. When different clock domains are used things become more complex. For example the counter register is manipulated over the bus interface and the re-scheduling of the Match event must take into account the delta between the two clock domains. The helper function below does the job and returns that delta. As an argument we pass the clock period of the other clock domain.

double
watch_dog::ScClockSyncTime(const double&
Period) {
  double NowTime =
    sc_simulation_time();
  double IntegerPart;
  if(modf(NowTime/Period, &IntegerPart)) {
    return(IntegerPart+1)*Period - NowTime;
  } else {//Clock domains are aligned
    return 0;
  }
}

The simple example above shows that the number of events can be reduced dramatically by using a different modeling approach. To give an idea how many behavioral events we saved we assume a simple use case and do not consider events created by accesses to the bus interface. The clock period does not change and the system is operating correctly and reloads the watchdog counter register in time so that the output port is not driven high. In that case not a single behavioral event has been fired (driving the output port high). The RTL style would have fired many events one for each decrement of the watchdog counter register. This will happen continuously after the watchdog counter register has been reloaded again. This modeling approach can be summarized by the following statement. “If there are no interactions with the model interface, there is no need to execute any model code”. The model is still timing accurate at the interface boundaries. The drawback is that we need a separate model for simulation and implementation. This approach can equally be used in the modeling of the internal state machine of any device model in that the “state” need only be updated when the device “interacts” with other elements in the simulation.

Clock Events

In general a pulsing clock causes a lot of events and easily reduces the simulation speed to the KIPS range. In the previous section the clock period has been communicated to model timing accurate behavior. One additional advantage of this style is that dynamic clock support comes for free. Whenever a clock generation unit communicates a new clock period the watchdog model connected to that clock signal can change its behavior accordingly. Again, changes of behavior need only be modeled if actions occur on the interface. SystemC per se does not support dynamic clocking as the period, duty cycles etc. are set during construction of the sc_clock object. It is obvious that we need dynamic clock support for today’s SoCs especially in the handheld market where battery life does matter. The function below demonstrates when the value of the clock period input port changes the device state can be updated. As the clock period has an impact on the Match event all future events are canceled and rescheduled. As a consequence of this change the current value of the watchdog counter register based on the previous clock period has to be calculated. The last action to take is to re-schedule the Match event based on the new watchdog timer period stored in mWatchDogClkPeriod.

watch_dog::WatchDogClkPeriodChanged() {
  MatchEvent.cancel();
  //Nothing to do when Watch Dog disabled
  if(WatchDogCtrl & 0x80) { //Enabled
  double NowTime = sc_simulation_time();
  //Calculate WatchDogValue
  WatchDogValue = WatchDogValue - (NowTime - mLastUpdate) / mWatchDogClkPeriod;

mWatchDogClkPeriod=WatchDogClkPeriod-
>read()); //Read new period from port
  if(mWatchDogClkPeriod) {
//When WatchDogClkPeriod is zero power down
   MatchEvent.notify(WatchDogValue *
     mWatchDogClkPeriod, SC_PS);
   }
  } else { //Only update mWatchDogClkPeriod
    mWatchDogClkPeriod = WatchDogClkPeriod-
>read();
  }
}

The callback function above to handle dynamic clocks also accepts a period of zero what means we completely switch off the device to save power (and conveniently simulation events). That’s all it takes to write an accurate and fast SystemC model that has little impact on simulation speed. Of course the coding methodology demonstrated here needs a different way of thinking compared to the RTL style often adopted by the hardware engineer.

3. SystemC Deficiencies

It has been stated that the SystemC sc_clock has no support for dynamic clocking but communicating the clock period via an input port overcomes that shortcoming. Unfortunately there is no kernel support to automatically re-schedule events when their time base changes. This behavior has to be implemented in each SystemC device and of course a base class providing that dynamic clock support is a convenient solution. It also becomes tricky when the model has several clock domains that have to be taken into account. Imagine a bridge between two buses clocked at different speeds. An incoming bus transaction at time n * tIN has to be re-transmitted at time m * tOUT
Where:
tIN is the clock period of the incoming bus,
tOUT is the clock period of the outgoing bus n and m are the number of clock cycles.

The model developer then manually has to calculate the timing delta to the next edge of the clock on the outgoing side. Again helper classes are extremely useful in that case in order to convert the clock domains.
Another deficiency is that there is no means to pass arguments to the SC_METHOD. This is helpful when you want to use a callback function to handle multiple instantiations of a sub-functionality. E.g. when you have an array of N timers in your model and you use a single callback function to handle all N timers. Via the argument you possibly could pass the number of the timer array you want to manipulate. Unfortunately the SystemC syntax does not support that. A minor issue is that SystemC does not warn you when you schedule an event that has already been scheduled. It will just pick the next one on the simulator event wheel. For debugging your model it would be helpful to get a warning.

4. Outcome

The discussed modeling methodology has been used to model a SystemC subsystem containing an interrupt controller, a UART and a timer model. That subsystem has been instantiated in a VaST Virtual System Platform with an ARM926e processor model and a fully arbitrating AHB bus (both models using VaST proprietary modeling methodology). The target software executing on the ARM model was embedded Linux and the simulation still ran at 50 MIPS. Not only the modeling style of the peripheral models is important with regards to simulation speed but also the other key components of a VSP. Those components are the simulator, the buses, bridges, memory and the Virtual Processor Models (VPMs). The VaST simulator can switch tasks efficiently within a few host machine cycles and the bus/bridge models typically ensure 1.5 million transactions per second. Nevertheless they are still cycle accurate on a transaction-level and fully arbitrating on a cycle basis. The VaST VPMs use well known techniques to run in the double digit MIPS range and ensure timing accuracy also. So although the SystemC peripherals are on a low bandwidth path in terms of events/access on the Linux embedded target code, it is seen that they contribute very little in terms of simulation performance. This would not be the case if they were modeled in a clocking RTL style.

5. Summary

The learned SystemC approach has proven to support accuracy and speed in a single model by reducing the number of events. Further recommendations are to avoid SC_THREADs when possible and use non-blocking interface method calls to prevent the user form using SC_THREADs. Those IMCs should support timing annotation as well as passing an event that could be fired by the slave when his bus transaction has been processed. On the bus interface PV modeling could be used in the first step to keep the modeling effort low but add timing details later on. With the use of annotated time the notion of scheduling future events can be followed. Following those guidelines enable the simulation of complex systems at reasonable speeds. Nevertheless SystemC does not have support for that methodology built in the simulation kernel. That’s why cumbersome calculations of event firing in the future are necessary.

6. References

[1] Groetker Thorsten, Liao Stan, Martin Grant, Swan Stuart: System Design with SystemC. 2002

[2] IEEE Standard SystemC® Language Reference Manual

[3] OSCI TLM 2.0 Draft 1