{{{
#!html
<h1>Writing TLM2.0-compliant timed SystemC simulation models for SoCLib</h1>
}}}

Authors : Alain Greiner, François Pêcheux, Aline Vieira de Mello
 
[[PageOutline]]

= A) Introduction =

This document is still under development.

It describes the modeling rules for writing TLM-T SystemC simulation models for SoCLib that are
compliant with the new TLM2.0 OSCI standard. 
These rules enforce the PDES (Parallel Discrete Event Simulation) principles. In the TLM-T approach, 
we don't use anymore the SystemC global time, as each PDES process involved in 
the simulation has its own local time. PDES processes (implemented as SC_THREADS) synchronize through 
messages piggybacked with time information. 
Models complying to these rules can be used with the "standard" OSCI simulation engine (SystemC 2.x) and
the TLM2.0 protocol, but can also be used also with others simulation engines, especially distributed, parallelized simulation engines.

The examples presented below use the VCI/OCP communication protocol selected by the SoCLib project,
but the TLM-T approach described here is very flexible, and is not limited to the VCI/OCP communication protocol.

The interested user should also look at the [WritingRules/General general SoCLib rules].

= B) Single VCI initiator and single VCI target =

Figure 1 presents a minimal system containing one single VCI initiator, '''my_initiator''' , and one single 
VCI target, '''my_target''' . The '''my_initiator''' module behavior is modeled by 
the SC_THREAD  '''execLoop()''', that contains an infinite loop. 
The call-back function '''vci_rsp_received()''' is executed when a VCI response packet is received by the initiator module.

[[Image(tlmt_figure_1.png, nolink)]]

Unlike the initiator, the target module has a purely reactive behaviour and is therefore modeled as a simple call-back function.
In other words, there is no need to use a SC_THREAD for this simple target component: the target behaviour is entirely described by the call-back 
function '''vci_cmd_received()''', that is executed when a VCI command packet is received by the target module.

The VCI communication channel is a point-to-point bi-directionnal channel, encapsulating two separated uni-directionnal 
channels: one to transmit the VCI command packet, one to transmit the VCI response packet.

= C) VCI initiator Modeling =

In the proposed example, the initiator module is modeled by the '''my_initiator''' class. 
This class inherits from the standard SystemC '''sc_core::sc_module''' class, that acts as the root class for all TLM-T modules.

The initiator local time is contained in a member variable named '''m_localTime''', of type '''sc_core::sc_time'''. The
local time can be accessed with the following accessors: '''addLocalTime()''', '''setLocalTime()''' 
and '''getLocalTime()'''.
{{{
  sc_core::sc_time m_localTime;                     // the initiator local time
  ...
  void addLocalTime(sc_core::sc_time t);            // add an increment to the local time
  void setLocalTime(sc_core::sc_time& t);           // set the local time
  sc_core::sc_time getLocalTime(void);              // get the local time
}}}

The initiator activity corresponds to the boolean member '''m_activity''' that indicates if the initiator is currently active.
This boolean variable is used by the arbitration threads i.e. '''true''', wants to participate in the arbitration in the interconnect) or inactive (i.e. '''false''', does not want
to participate in the arbitration in the interconnect). The corresponding access functions are '''setActivity()''' and '''getActivity()'''.
{{{
  bool m_activity;
  ...
  void setActivity(bool t);                         // set the activity status (true if the component is active)
  bool getActivity(void);                           // get the activity state
}}}

The '''execLoop()''' method, describing the initiator behaviour must be declared as a member function of 
the '''my_initiator''' class. 

Finally, the class '''my_initiator''' must contain a member variable '''p_vci_init''', of type '''tlmt_simple_initiator_socket'''. 
This member variable represents the VCI initiator port. It has 3 template parameters, two of which are used to help 
connecting the response callback function ('''my_initiator''' in the example, first template parameter) to the port and 
defining the port type
('''soclib_vci_types''' in the following example, third template parameter). '''soclib_vci_types''' is indeed a C++ structure containing two
typedef: the first typedef defines the payload type as VCI, and the other defines the TLM phase type. The phase type can either
be '''TLMT_CMD''' (i.e. the transaction indicates the emission of a command by an initiator and its reception by a target),
'''TLMT_RSP''' (i.e. the transaction indicates the emission of a response by a target and its reception by an initiator),
or '''TLMT_INFO''' (i.e. a TLM-T transaction emitted by one side of a link (vci, irq or fifo) to get information such as time
and activity on the other side of the link).

== C.1) Sending a VCI command packet ==

To send a VCI command packet, the '''execLoop()''' method must use the '''nb_transport_fw()''' method, that is a member 
function of the '''p_vci_init''' port. The prototype of this method is the following:
{{{

  tlm::tlm_sync_enum nb_transport_fw               /// sync status
  ( soclib_vci_types::tlm_payload_type &payload,      ///< VCI payload pointer
    soclib_vci_types::tlm_phase_type   &phase,       ///< transaction phase
    sc_core::sc_time                   &time);        ///< time
}}}

The first parameter of this member function is the VCI packet, the second represents the phase (TLMT_CMD in this case), and the third
parameter contains the initiator local time.

To prepare a VCI packet for sending, the '''execLoop''' function must declare two objects locally, '''payload''' and '''phase'''.
{{{
  soclib_vci_types::tlm_payload_type payload;
  soclib_vci_types::tlm_phase_type phase;
}}}
A payload of type '''soclib_vci_types::tlm_payload_type''' corresponds to a '''tlmt_vci_transaction'''.
It contains three groups of information: 
 * TLM2.0 related fields
 * TLM-T related fields 
 * VCI related fields 
The contents of a '''tlmt_vci_transaction''' is defined below:
{{{
class tlmt_vci_transaction
{
	...
private:
  // TLM2.0 related fields and common structure

  sc_dt::uint64        m_address; 		// address
  unsigned char*       m_data; 			// buf
  unsigned int         m_length; 		// nword
  tlmt_response_status m_response_status; 	// rerror
  bool                 m_dmi; 			// nothing
  unsigned char*       m_byte_enable; 		// be
  unsigned int         m_byte_enable_length; 
  unsigned int         m_streaming_width; 	// 

  // TLM-T related fields
  
  bool*                m_activity_ptr;
  sc_core::sc_time*    m_local_time_ptr;

  // VCI related fields
  
  tlmt_command         m_command; 		// cmd
  unsigned int         m_src_id; 		// srcid
  unsigned int         m_trd_id; 		// trdid
  unsigned int         m_pkt_id; 		// pktid
}}}

The TLM2.0 compliant accessors allow to set the TLM2.0 related fields, such as the transaction address, the byte enable
array pointer and its associated size in bytes, and the data array pointer and its associated size in bytes. The byte
enable array allows to build versatile packets thanks to a powerful but slow data masking scheme. Further experiments are
currently done to evaluate the performance degradation incurred by the byte formatting.
It is therefore  possible that the types of the '''m_data''' and '''m_byte_enable''' of the '''tlmt_vci_transaction'''
will be changed to '''uint32*''' in a near future.

Dedicated VCI accessors are used to define the VCI transaction type, that can either be '''set_read()''' (for read command),
'''set_write()''' (for write command),'''set_locked_read()''' (for atomic locked read), 
and '''set_store_cond()''' (for atomic store conditional). The '''set_src_id()''', '''set_trd_id()''' and '''set_pkt_id()''' functions 
respectively set the VCI source, thread and packet identifiers.
The following example describes a VCI write command:
{{{
    payload.set_address(0x10000000);//ram 0
    payload.set_byte_enable_ptr(byte_enable);
    payload.set_byte_enable_length(nbytes);
    payload.set_data_ptr(data);
    payload.set_data_length(nbytes); // 5 words of 32 bits

    payload.set_write();
    payload.set_src_id(m_id);
    payload.set_trd_id(0);
    payload.set_pkt_id(pktid);

    phase= soclib::tlmt::TLMT_CMD;
    sendTime = getLocalTime();

    p_vci_init->nb_transport_fw(payload, phase, sendTime);
}}}


The '''nb_transport_fw()''' function is non-blocking. 
To implement a blocking transaction (such as a cache line read, where the processor is stalled during the VCI transaction), 
the model designer must use the SystemC '''sc_core::wait(x)''' primitive ('''x''' being of type '''sc_core::sc_event'''): 
the '''execLoop()''' thread is then suspended, and will be reactivated when the response packet is actually received.

== C.2) Receiving a VCI response packet ==

To receive a VCI response packet, a call-back function must be defined as a member function of the 
class '''my_initiator'''. This call-back function (named '''vci_rsp_received()''' in the example), must be
declared in the '''my_initiator''' class and
is executed each time a VCI response packet is received on the '''p_vci_init''' port. The function name is not 
constrained, but the arguments must respect the following prototype:
{{{
  tlm::tlm_sync_enum vci_rsp_received              
    ( soclib_vci_types::tlm_payload_type &payload,     // payload
      soclib_vci_types::tlm_phase_type   &phase,      // transaction phase
      sc_core::sc_time                   &time);       // resp time
}}}
The return value (type tlm::tlm_sync_enum)  must be sytematically set to tlm::TLM_COMPLETED in this implementation
The function parameters are identical to those described in the forward transport function 

In the general case, the actions executed by the call-back function depend on the response transaction type ('''m_command''' field), as well as 
the '''pktid''' and '''trdid''' fields.
For sake of simplicity, the call-back function proposed below does not make any distinction between VCI transaction types.

== C.3) Initiator Constructor ==

The constructor of the class '''my_initiator''' must initialize all the member variables, including 
the '''p_vci_init''' port. The '''vci_rsp_received()''' call-back function being executed in the context of the thread sending 
the response packet, a link between the '''p_vci_init''' port and this call-back function must be established. 

The '''my_initiator''' constructor for the '''p_vci_init''' object must be called with the following arguments:
{{{
  p_vci_init.register_nb_transport_bw(this, &my_initiator::vci_rsp_received);
}}}

== C.4) Lookahead parameter ==

The SystemC simulation engine behaves as a cooperative, non-preemptive multi-tasks system. Any thread in the system must stop execution 
after at some point, in order to allow the other threads to execute. With the proposed approach, a TLM-T initiator will never stop if 
it does not execute blocking communication (such as a processor that has all code and data in the L1 caches). 

To solve this issue, it is necessary to define -for each initiator module- a '''lookahead''' parameter. This parameter defines the maximum 
number of cycles that can be executed by the thread before it is automatically stopped. The '''lookahead''' parameter allows the system designer 
to bound the de-synchronization time interval between threads. 

A small value for this parameter results in a better timing accuracy for the simulation, but implies a larger number of context switches, 
and a slower simulation speed. 

== C.4) VCI initiator example ==

{{{

////////////////////////// my_initiator.h ////////////////////////////////


////////////////////////// my_initiator.cpp ////////////////////////////////


}}}

= D) VCI target modeling =

In the proposed example, the '''my_target''' component handles all VCI commands in the same way, and there is no error management.

The class '''my_target''' inherits from the class '''sc_core::sc_module'''. The class '''my_target''' contains a member 
variable '''p_vci_target''' of type '''tlmt_simple_target_socket'''. This object has 3 template parameters, that are identical to
those used for declaring initiator ports (see above).

== D.1) Receiving a VCI command packet ==

To receive a VCI command packet, a call-back function must be defined as a member function of the class '''my_target'''. 
This call-back function (named '''vci_cmd_received()''' in the example), will be executed each time a VCI command packet is received on 
the '''p_vci_target''' port. The function name is not constrained, but the arguments must respect the following prototype:
{{{
  tlm::tlm_sync_enum vci_cmd_received              
  ( soclib_vci_types::tlm_payload_type &payload,      // VCI payload pointer
    soclib_vci_types::tlm_phase_type   &phase,       // transaction phase
    sc_core::sc_time                   &time);        // time
}}}  

== D.2) Sending a VCI response packet ==

To send a VCI response packet the call-back function '''vci_cmd_received()''' must use the '''nb_transport_bw()''' method, that is a member function of 
the class '''tlmt_simple_target_socket''', and has the same arguments as the '''nb_transport_fw()''' function.
Respecting the general TLM2.0 policy, the payload argument refers to the same '''tlmt_vci_transaction''' object for both the '''nb_transport_fw()''' and '''nb_transport_bw()''' functions,
and the associated call-back functions. The set_response_status field must be documented for all transaction types, but only two values are used in this implementation: 
 * TLMT_OK_RESPONSE
 * TLMT_ERROR_RESPONSE
For a reactive target, the response packet time is computed as the command packet time plus the target intrinsic latency.

{{{
	payload.set_response_status(soclib::tlmt::TLMT_OK_RESPONSE);
	phase = soclib::tlmt::TLMT_RSP;
	time = time + (nwords * UNIT_TIME);
	p_vci_target->nb_transport_bw(payload, phase, time);
}}}

== D.3) Target Constructor ==

The constructor of the class '''my_target''' must initialize all the member variables, including 
the '''p_vci_target''' port. The '''vci_cmd_received()''' call-back function being executed in the context of the thread sending 
the command packet, a link between the '''p_vci_target''' port and the call-back function must be established. 
The '''my_target''' constructor must be called with the following arguments:
{{{
  p_vci_target.register_nb_transport_fw(this, &my_target::vci_cmd_received);
}}}

== D.4) VCI target example ==
 {{{

////////////////////////// my_target.h ////////////////////////////////


////////////////////////// my_target.cpp ////////////////////////////////


}}}

= E) VCI Interconnect modelling =

The VCI interconnect used for the TLM-T simulation is a generic simulation model, named '''!VciVgmn'''.
The two main parameters are the number of initiators, and the number of targets. In TLM-T simulation,
we don't want to reproduce the cycle-accurate behavior of a particular interconnect. We only want to simulate the contention in 
the network, when several VCI intitiators try to reach the same VCI target.
Therefore, the network is actually modeled as a complete cross-bar : In a physical network such as the multi-stage network described 
in Figure 2.a, conflicts can appear at any intermediate switch. In the '''!VciVgmn''' network described in Figure 2.b, conflicts can 
only happen at the output ports. It is possible to specify a specific latency for each input/output couple. As in most physical 
interconnects, the general arbitration policy is round-robin.

[[Image(tlmt_figure_2.png, nolink)]]

== E.1) Generic network modeling ==

 There is actually two fully independent networks for VCI command packets and VCI response packets. There is a routing function for each input 
 port, and an arbitration function for each output port, but the two networks are not symmetrical :
 * For the command network, the arbitration policy is distributed: there is one arbitration thread for each output port 
 (i.e. one arbitration thread for each VCI target). Each arbitration thread is modeled by a SC_THREAD, and contains a local clock.
 * For the response network, there are no conflicts, and there is no need for arbitration. Therefore, there is no thread 
 (and no local time) and the response network is implemented by simple function calls.

This is illustrated in Figure 3 for a network with 2 initiators and three targets :  

[[Image(tlmt_figure_3.png, nolink)]]

== E.2) Arbitration Policy ==

As described above, there is one '''cmd_arbitration''' thread associated to each VCI target. This thread is in charge of selecting one timed request between
all possible requesters, and to forward it to the target. According to the PDES principles, the arbitration thread must select the request with the smallest timestamp. 
The arbitration process must take into account the actual state of the VCI initiators: For example a DMA coprocessor that has not yet been activated
will not send request and should not participate in the arbitration process. As a general rule, each VCI initiator must define an '''active''' boolean flag,
defining if it should participate to the arbitration. This '''active''' flag is always set to true for general purpose processors.
Any arbitration thread receiving a timed request is resumed. It must obtain an up to date timing & activity information for all its input channels before making any decision.
To do that, the LocalTime and ActivityStatus of all VCI initiators are considered as global variables, that can be accessed (for read only) by all arbitration threads.  
The arbitration policy is the following : The arbitration thread scans all its input channels, and selects the smallest time between the active initiators.
If there is a request, this request is forwarded to the target, and the arbitration thread local time is updated. 
If not, the thread is descheduled and will be resumed when it receives a new request.

For efficiency reasons, in this implementation, each arbitration thread constructs - during elaboration of the simulation - two local array of pointers (indexed
by the input channel index) to access the LocalTime and ActivityStatus variables of the corresponding VCI initiators. To get this information, each arbitration thread 
uses the ''nb_transport_bw()''' function on all its VCI target ports, with a a dedicated phase called '''soclib::tlmt::TLMT_INFO'''. The payload argument refers to the same 
'''tlmt_vci_transaction''' object as the two other phases (TLMT_CMD and TLMT_RSP).

{{{
for (size_t i=0;i<m_nbinit;i++) {
    phase   = soclib::tlmt::TLMT_INFO;
    m_RspArbCmdRout[i]->p_vci->nb_transport_bw(payload, phase, rspTime);
    m_array[i].activity = payload.get_activity_ptr();
    m_array[i].time = payload.get_local_time_ptr();	
}
}}}

As the net-list of the simulated pltform mus be explicitely defined before constructing those LocalTime and ActivityStatus arrays, the vgmn hardware component 
provides an utility function '''fill_time_activity_arrays()''' that must be called in the SystemC top-cell, before starting the simulation.