mirror of
https://github.com/nasa/trick.git
synced 2024-12-20 13:43:10 +00:00
784 lines
34 KiB
Markdown
784 lines
34 KiB
Markdown
|
The Simulation Definition File or S_define is the file which lays out the architecture
|
|||
|
of the simulation. Details about job frequencies, job class, job data, importing/exporting
|
|||
|
data to other simulations, freeze cycles, integration, etc. are all housed in this one file.
|
|||
|
It is the file which Trick's main processor, CP, parses to determine what will be part of
|
|||
|
the simulation and how the pieces fit together.
|
|||
|
|
|||
|
This section begins with the syntax for the S_define, then details each and every entry
|
|||
|
in the syntax template. This section also details how to use Trick's main processor, CP.
|
|||
|
|
|||
|
```
|
|||
|
[/* [TRICK_HEADER]
|
|||
|
|
|||
|
PURPOSE: (purpose statement)
|
|||
|
[LIBRARY DEPENDENCIES:
|
|||
|
(
|
|||
|
[(<rel_path_of_model>/<model_source_code_n.c>)]
|
|||
|
)
|
|||
|
]
|
|||
|
|
|||
|
[DEFAULT_DATA:
|
|||
|
(
|
|||
|
[(struct_type instance_name <rel_path_of_model>/<default_data_file>)]
|
|||
|
)
|
|||
|
]
|
|||
|
*/]
|
|||
|
|
|||
|
#include "sim_objects/default_trick_sys.sm"
|
|||
|
|
|||
|
##include "<rel_path_of_model>/<model_header_file.h>"
|
|||
|
|
|||
|
%header{
|
|||
|
/* User header code block */
|
|||
|
%}
|
|||
|
|
|||
|
%{
|
|||
|
/* User code block */
|
|||
|
%}
|
|||
|
|
|||
|
class <sim_object_type_name> : public Trick::SimObject {
|
|||
|
|
|||
|
[(public|protected|private):]
|
|||
|
|
|||
|
<CLASS_TYPE|STRUCTURE_TYPE|ENUMERATED_TYPE|intrinsic_type> [*]* <data_varaible_name> [dims]*
|
|||
|
|
|||
|
<sim_object_type_name>([args]) {
|
|||
|
[C<#>] [{job_tag}] [P<#>] ([<cycle_time>, [<start_time>, [<stop_time>,]]] <job_class>) \
|
|||
|
[<return_var> =] <module>([args]) ;
|
|||
|
|
|||
|
}
|
|||
|
|
|||
|
[<method_name>([args]) { ... } ;]
|
|||
|
} ;
|
|||
|
|
|||
|
[job_class_order {
|
|||
|
<job_class_1>,
|
|||
|
<job_class_2>,
|
|||
|
...
|
|||
|
} ;]
|
|||
|
|
|||
|
[<sim_object_type_name> <sim_object_instantiation>[(args)] ;]
|
|||
|
|
|||
|
[integrate <integrator_name> (<integration_dt>) <sim_object_name> [,<sim_object_name] ;]
|
|||
|
|
|||
|
[collect <reference> = {[<reference> [,<reference>]]};]
|
|||
|
|
|||
|
[void create_connections() { <glue_code> }]
|
|||
|
```
|
|||
|
|
|||
|
### Trick Header Comment
|
|||
|
|
|||
|
This optional section of the S_define file is a C style comment found anywhere in the S_define file.
|
|||
|
CP will parse the Trick Header comment looking for library dependencies and default data. Library
|
|||
|
dependencies are model source code files that are added to the list of files to be compiled and
|
|||
|
linked into the simulation. These dependencies differ from the ones found in the actual model source
|
|||
|
code in that they are the full relative path to the actual source code file, not the resulting object file.
|
|||
|
CP also looks for old style default data files. Each default data entry has 3 fields, the structure type, the
|
|||
|
instance name, and the relative path to the default data file. CP will read in the default data file
|
|||
|
substituting each occurrence of the structure type in the file with the instance name. All of the default
|
|||
|
data files are concatenated to the S_default.dat file.
|
|||
|
|
|||
|
#### S_define Library Dependencies
|
|||
|
```
|
|||
|
LIBRARY DEPENDENCY:
|
|||
|
((relative_path/model_1.c)
|
|||
|
(relative_path/model_2.cpp))
|
|||
|
```
|
|||
|
|
|||
|
Library dependencies list out model source code files required by the simulation. There are several
|
|||
|
locations to add library dependencies, one of which is in the S_define file. The format of
|
|||
|
dependencies in the S_define file is a relative path to the model source file. The path is relative
|
|||
|
to -I include paths found in TRICK_CFLAGS and TRICK_CXXFLAGS.
|
|||
|
|
|||
|
For example if the full path to our model is /this/is/a/full/path/to/model.c and in our TRICK_CFLAGS
|
|||
|
we have -I/this/is/a/full as one of the included search paths, the library dependency must complete the
|
|||
|
full path to the file, in this case path/to/model.c. Library dependendencies in the S_define file
|
|||
|
differ from ones found in model source code as they must be the full path to the source file not the
|
|||
|
object file.
|
|||
|
|
|||
|
### Include files
|
|||
|
|
|||
|
There are two types of includes in the S_define file.
|
|||
|
|
|||
|
#### Single pound "#" includes.
|
|||
|
|
|||
|
Include files with a single pound "#" are parsed as they are part of the S_define file. They are
|
|||
|
treated just as #include files in C or C++ files. These files usually include other sim objects or
|
|||
|
instantiations as part of the S_define file.
|
|||
|
|
|||
|
#### Double pound "#" includes.
|
|||
|
|
|||
|
Include files with a double pound "##" are not parsed as part of the S_define file. These files are the
|
|||
|
model header files. They include the model class and structure definitions as well as C prototypes for
|
|||
|
functions used in the S_define file. Double pound files are copied, minus one pound, to S_source.hh.
|
|||
|
|
|||
|
### User Header Code Block
|
|||
|
|
|||
|
This section of the S_define (encapsulated by "%header{...%}") can be used for including header files
|
|||
|
directly into the S_source.hh. Header files listed here will not be input processed.
|
|||
|
|
|||
|
### User Code Block
|
|||
|
|
|||
|
This section of the S_define (encapsulated by %{.....%}) can be used for any user specific
|
|||
|
global C code. CP will simply insert this section of code into the S_source.c file after
|
|||
|
all header files are included. Typically this feature is used as a quick method for customizing
|
|||
|
simulations with additions of global variables and functions.
|
|||
|
|
|||
|
### Simulation Object Definition
|
|||
|
|
|||
|
A simulation definition file may contain any number of simulation object definitions.
|
|||
|
A simulation object definition is of the form: class <sim_object_name> : public Trick::SimObject { ... }.
|
|||
|
All sim objects must inherit from the Trick::SimObject or a derivative. A sim object definition
|
|||
|
may contain zero or more data structure declarations and zero or more module declarations.
|
|||
|
|
|||
|
### Model Classes and Data Structures
|
|||
|
|
|||
|
Model classes and data structures are declared within a sim object. Model classes and data structures
|
|||
|
may be public, protected, or private within the sim object. Standard C++ privacy rules apply to
|
|||
|
all data with the sim object. Sim object protected and private data will not be accessible to the input
|
|||
|
processor.
|
|||
|
|
|||
|
Intrinsic types are allowed as sim object data members.
|
|||
|
|
|||
|
### Job Declarations
|
|||
|
|
|||
|
Jobs are the hands and feet of the simulation. They are the building blocks for the
|
|||
|
simulation. Jobs are C or C++ routines with special Trick tags that determine scheduling,
|
|||
|
object dependencies, etc.
|
|||
|
|
|||
|
Jobs only appear in the constructor of the sim object.
|
|||
|
|
|||
|
```
|
|||
|
[C<#>] [{job_tag}] [P<#>] ([<cycle_time>, [<start_time>, [<stop_time>,]]] <job_class>) <module>([args]) ;
|
|||
|
```
|
|||
|
|
|||
|
Most of these fields are optional depending on how the module is classified or utilized
|
|||
|
within the sim. The following subsections detail the usage and purpose of each of these fields.
|
|||
|
|
|||
|
#### Child Thread Specification
|
|||
|
|
|||
|
The first field of a module declaration is an optional child process specification in
|
|||
|
the form of a capital C immediately followed by an integer child ID number; i.e. C1, C2, C3,
|
|||
|
etc. The child process specification allows parallel processing for simulation modules.
|
|||
|
|
|||
|
Every simulation has a parent process. A child specification will result in the spawning of
|
|||
|
a thread for each distinct child ID specified. More than one job can be specified for each child ID.
|
|||
|
Jobs with child specifications will run in parallel with other jobs within each software frame,
|
|||
|
so users may be required to specify job dependencies (see Section 4.4.10) to keep parallel jobs
|
|||
|
from stepping on each other's common data access. The collection of the parent process and all
|
|||
|
of its children defined in one S_define file is referred to as a Process Group (PG). A simulation
|
|||
|
can be composed of multiple synchronized or non-synchronized PGs which will be discussed in more
|
|||
|
detail in Section 4.4.12. and Section 7.2.
|
|||
|
|
|||
|
In most cases, for maximum execution performance, jobs should only be specified to run on a child
|
|||
|
process if the host workstation has more than one CPU; and preferably, one CPU per child specified.
|
|||
|
With the same rule in mind, using more than one PG in the simulation design should only occur when
|
|||
|
the simulation has parallel components and multiple process/multiple computers are available. When
|
|||
|
the host workstation has only one CPU, a simulation with a job running on a child will be much slower
|
|||
|
than the same simulation with no children. There are exceptions to this rule, like running asynchronous
|
|||
|
background modules on a child that only runs the child during the wait cycle of a simulation set up for
|
|||
|
real-time operations.
|
|||
|
|
|||
|
Child IDs start at 1 and may skip any integer values. The Trick Executive will spawn enough threads
|
|||
|
to accomodate the highest Child ID specified in the S_define file. Jobs without a child specification
|
|||
|
will be executed by the parent process. Jobs with a C1 child specification will be executed by the
|
|||
|
first child thread; jobs with a C2 specification will be executed by the second child thread; and so on.
|
|||
|
|
|||
|
Child Threads have three different scheduling choices. See Section XYZ for child thread scheduling
|
|||
|
details.
|
|||
|
|
|||
|
#### Job Tagging
|
|||
|
|
|||
|
This optional field allows you to tag a job or set of jobs. The tag is surrounded in curly
|
|||
|
braces. In the input file, you may then operate on the tag. All jobs with the same tag will be
|
|||
|
modified in the same manner. For example, if jobA and jobB are tagged BLUE, the input file may
|
|||
|
have a statement:
|
|||
|
|
|||
|
```python
|
|||
|
trick.add_read(13.0, """trick.exec_set_job_cycle("BLUE" , 0.001)""")
|
|||
|
```
|
|||
|
|
|||
|
This will change the frequency of the jobs to 1 millisecond. You might also disable the jobs
|
|||
|
tagged BLUE with the following:
|
|||
|
|
|||
|
```python
|
|||
|
trick.add_read(10.0, """trick.exec_set_job_onoff("BLUE" , False)""")
|
|||
|
```
|
|||
|
|
|||
|
#### Job Phasing
|
|||
|
|
|||
|
The next field of a job declaration is an optional phase number specification in the form of a
|
|||
|
capital P immediately followed by an integer phase ID number from 1 to 65534, e.g., P1, P2, P3, etc.
|
|||
|
Without a specified phase field, the default phase number is 60000. The phase specification may be
|
|||
|
used on all class jobs to sequence the execution of jobs in the same class. Jobs tagged with P1
|
|||
|
execute first, then P2, etc. Jobs with the same phase number are executed in the order they are
|
|||
|
in the S_define.
|
|||
|
|
|||
|
#### Execution Schedule Time Specification
|
|||
|
|
|||
|
The execution schedule specification specifies the job's execution cycle time, start time, and
|
|||
|
stop time. The time specs must be a comma separated list of floating point numbers enclosed by
|
|||
|
parentheses, e.g. (1.0,0.5,10.0) would execute a module starting at 0.5 seconds, and every 1.0
|
|||
|
seconds thereafter until 10.0 seconds was reached (9.5 seconds would be the time of the last
|
|||
|
job call). The start and stop fields are optional; e.g. (1.0,0.5) does the same as the previous
|
|||
|
example except the module will keep being called after 10.0 seconds. Also, a (1.0) specification
|
|||
|
will start the job at 0.0 (the default) and continue calling the job at 1.0 second intervals.
|
|||
|
|
|||
|
Only the jobs categorized as CYCLIC or also
|
|||
|
the freeze_scheduled job class (see Table SD_1 below) require the execution time specification.
|
|||
|
All schedule time specifications are in seconds.
|
|||
|
|
|||
|
All other job classes categorized in Table SD_1 should NOT specify an execution time specification:
|
|||
|
- NON-CYCLIC (red color) of course do not require a spec because they run only once.
|
|||
|
- FRAME-BOUNDARY (blue color) are tied to each execution frame and not a cycle time.
|
|||
|
- INTEGRATOR (cyan color) are specified via the state integration specifications for the simulation (see Section 4.4.7).ME-BOUNDARY (blue color) are tied to each execution frame and not a cycle time.
|
|||
|
- INTEGRATOR (cyan color) are specified via the state integration specifications for the simulation (see Section 4.4.7).
|
|||
|
- SELF-SCHEDULING (yellow color) are responsible for scheduling themselves.
|
|||
|
- FREEZE (purple color) are tied to the freeze frame (EXCEPT for freeze_scheduled class jobs which DO need the spec.)
|
|||
|
- CHECKPOINT (orange color) are tied to checkpointing and run only once.
|
|||
|
|
|||
|
##### Job Class
|
|||
|
The job class determines where in the simulation a job is run. Each job in the S_define file
|
|||
|
requires a job class.
|
|||
|
|
|||
|
<center>
|
|||
|
<b>KEY: Job Class Category Colors used in Table SD_1</b>
|
|||
|
<table>
|
|||
|
<tr bgcolor="#c0c0c0"><th>Category Color</th><th>The job classes in Table SD_1 are categorized as one of the following:</th></tr>
|
|||
|
<tr><td bgcolor="#ff8080">NON-CYCLIC</td><td>jobs that run once either at simulation startup or termination</td></tr>
|
|||
|
<tr><td bgcolor="#add8e6">FRAME-BOUNDARY</td><td>jobs that run cyclically before/after the scheduled job loop</td></tr>
|
|||
|
<tr><td bgcolor="#ccffff">INTEGRATOR</td><td>jobs that run first in the scheduled job loop for Trick state integration</td></tr>
|
|||
|
<tr><td bgcolor="#ffff99">SELF-SCHEDULING</td><td>jobs in the scheduled job loop that must schedule themselves when to run</td></tr>
|
|||
|
<tr><td bgcolor="#ccffcc">CYCLIC</td><td>jobs in the scheduled job loop that run cyclically according to their specified cycle time</td></tr>
|
|||
|
<tr><td bgcolor="#e6ceff">FREEZE</td><td>jobs that are run during the freeze execution mode</td></tr>
|
|||
|
<tr><td bgcolor="#ffc285">CHECKPOINT</td><td>jobs that are run when a checkpoint is dumpded or loaded</td></tr>
|
|||
|
</table>
|
|||
|
|
|||
|
<b>Table SD_1 Trick-Provided Job Classes</b>
|
|||
|
<table>
|
|||
|
<tr bgcolor="#c0c0c0"><th>Job Class Name</th><th>Description</th></tr>
|
|||
|
<tr><td bgcolor="#ff8080">default_data</td><td>Module executes only once at simulation time = zero.
|
|||
|
Called before input file is read.</td></tr>
|
|||
|
<tr><td bgcolor="#ff8080">initialization</td><td>Module executes only once at simulation time = zero.
|
|||
|
Called after input file is read.</td></tr>
|
|||
|
<tr bgcolor="add8e6"><td colspan="2" align="center">BEGIN EXECUTION FRAME</td></tr>
|
|||
|
<tr><td bgcolor="add8e6">top_of_frame</td><td>Module runs at the beginning of every execution frame, before the scheduled job loop,
|
|||
|
even before any child threads are started for the frame.</td></tr>
|
|||
|
<tr bgcolor="ccffcc"><td colspan="2" align="center">BEGIN SCHEDULED JOB LOOP</td></tr>
|
|||
|
<tr><td bgcolor="#ccffff">pre_integration</td><td>Runs only once before the integration loop.
|
|||
|
For example, in the case of a Runge Kutta 4, the derivative and integration jobs will be called
|
|||
|
four times during the integration steps. A pre_integration job will execute a
|
|||
|
single time before this four step integration process has occurred.</td></tr>
|
|||
|
<tr bgcolor="ccffff"><td colspan="2" align="center">BEGIN INTEGRATION LOOP</td></tr>
|
|||
|
<tr><td bgcolor="#ccffff">derivative</td><td>Equations of motion (EOM) derivative function.</td></tr>
|
|||
|
<tr><td bgcolor="#ccffff">integration</td><td>Equations of motion state integration function.</td></tr>
|
|||
|
<tr bgcolor="ccffff"><td colspan="2" align="center">END INTEGRATION LOOP</td></tr>
|
|||
|
<tr><td bgcolor="#ccffff">post_integration</td><td>Runs only once after the integration loop.</td></tr>
|
|||
|
<tr><td bgcolor="#ccffff">dynamic_event</td><td>Runs after the integration loop
|
|||
|
(after any post_integration jobs). Provides a continuous time dependent equation whose root
|
|||
|
defines a discontinuous event in the system EOM, evaluation of function returns an
|
|||
|
estimated delta time to reach the root.</td></tr>
|
|||
|
<tr><td bgcolor="ffff99">automatic</td><td>Module which reschedules its own next call time, and will run before other CYCLIC jobs in the frame.</td></tr>
|
|||
|
<tr><td bgcolor="#ccffcc">environment</td><td>Time dependent boundary conditions (mother nature).</td></tr>
|
|||
|
<tr><td bgcolor="#ccffcc">sensor</td><td> Simulated interface between dynamics and control simulation components.</td></tr>
|
|||
|
<tr><td bgcolor="#ccffcc">sensor_emitter</td><td>Same as sensor, but for the emitter portion of an integrated
|
|||
|
multi-component sensor system.</td></tr>
|
|||
|
<tr><td bgcolor="#ccffcc">sensor_reflector</td><td>Same as sensor, but for the reflector portion of an
|
|||
|
integrated multi-component sensor system.</td></tr>
|
|||
|
<tr><td bgcolor="#ccffcc">sensor_receiver</td><td>Same as sensor, but for the receiver portion of an
|
|||
|
integrated multi-component sensor system.</td></tr>
|
|||
|
<tr><td bgcolor="#ccffcc">scheduled</td><td>Typical flight software and hardware subsystems.</td></tr>
|
|||
|
<tr><td bgcolor="#ccffcc">effector</td><td>Simulated interface between control and force generator
|
|||
|
simulation components.</td></tr>
|
|||
|
<tr><td bgcolor="#ccffcc">effector_emitter</td><td>Same as effector, but for the action portion of an
|
|||
|
action-reaction effector system.</td></tr>
|
|||
|
<tr><td bgcolor="#ccffcc">effector_receiver</td><td>Same as effector, but for the reaction portion of an
|
|||
|
action-reaction effector system.</td></tr>
|
|||
|
<tr><td bgcolor="#ccffcc">logging</td><td>Simulation data recording or displaying.</td></tr>
|
|||
|
<tr><td bgcolor="#ffff99">automatic_last</td><td>Module which reschedules its own next call time, and will run after other CYCLIC jobs in the frame.</td></tr>
|
|||
|
<tr bgcolor="ccffcc"><td colspan="2" align="center">END SCHEDULED JOB LOOP</td></tr>
|
|||
|
<tr><td bgcolor="add8e6">end_of_frame</td><td>Module runs at the end of every execution frame, after the scheduled job loop completes.</td></tr>
|
|||
|
<tr bgcolor="add8e6"><td colspan="2" align="center">END EXECUTION FRAME</td></tr>
|
|||
|
<tr><td bgcolor="#e6ceff">freeze_init</td><td>Module executes only once upon entering FREEZE mode.</td></tr>
|
|||
|
<tr bgcolor="e6ceff"><td colspan="2" align="center">BEGIN FREEZE FRAME</td></tr>
|
|||
|
<tr bgcolor="e6ceff"><td colspan="2" align="center">BEGIN FREEZE LOOP</td></tr>
|
|||
|
<tr><td bgcolor="#e6ceff">freeze_scheduled</td><td>Module executes cyclically during simulation FREEZE mode according to its specified cycle time.</td></tr>
|
|||
|
<tr bgcolor="e6ceff"><td colspan="2" align="center">END FREEZE LOOP</td></tr>
|
|||
|
<tr><td bgcolor="#e6ceff">freeze</td><td>Module runs at end of every freeze frame, after the freeze loop completes.</td></tr>
|
|||
|
<tr bgcolor="e6ceff"><td colspan="2" align="center">END FREEZE FRAME</td></tr>
|
|||
|
<tr><td bgcolor="#e6ceff">unfreeze</td><td>Module executes only once upon leaving FREEZE mode before returning to RUN mode.</td></tr>
|
|||
|
<tr><td bgcolor="#ffc285">checkpoint</td><td>Module executes only once before a checkpoint is dumped.</td></tr>
|
|||
|
<tr><td bgcolor="#ffc285">post_checkpoint</td><td>Module executes only once after a checkpoint is dumped.</td></tr>
|
|||
|
<tr><td bgcolor="#ffc285">preload_checkpoint</td><td>Module executes only once before a checkpoint is loaded.</td></tr>
|
|||
|
<tr><td bgcolor="#ffc285">restart</td><td>Module executes only once after a checkpoint is loaded.</td></tr>
|
|||
|
<tr><td bgcolor="#ff8080">shutdown</td><td>Module executes only once at simulation termination to allow a
|
|||
|
graceful shutdown.</td></tr>
|
|||
|
</table>
|
|||
|
</center>
|
|||
|
|
|||
|
##### Job Return Type
|
|||
|
|
|||
|
All integration class jobs must return an integer value which represents the current integration
|
|||
|
pass identifier. If all integration passes are complete, the job must return a zero.
|
|||
|
|
|||
|
All dynamic_event class jobs must return a double precision value representing the estimated time
|
|||
|
to go to the defined event, in seconds.
|
|||
|
|
|||
|
All other jobs managed by the Trick executive, not integration or dynamic_event, may return any
|
|||
|
type. If a function is declared with an integer return value, the job must return a non-negative
|
|||
|
integer or else the executive will assume an error occurred an immediately terminate the simulation.
|
|||
|
If the job does not return an integer, Trick will not take any action based on a return value. Note
|
|||
|
that initialization job return values are NOT checked by the executive.
|
|||
|
|
|||
|
##### Job Name
|
|||
|
|
|||
|
This field specifies the job name of the job as defined in the job’s source code.
|
|||
|
|
|||
|
C function and C++ member functions can be called as a jobs. Here is a quick example of a C and C++
|
|||
|
calls.
|
|||
|
|
|||
|
```C++
|
|||
|
%{
|
|||
|
extern "C" {
|
|||
|
c_function() ;
|
|||
|
}
|
|||
|
%}
|
|||
|
|
|||
|
class MySimObject() : public Trick::SimObject {
|
|||
|
|
|||
|
public:
|
|||
|
Ball my_ball ;
|
|||
|
|
|||
|
MySimObject() {
|
|||
|
(1.0 , "scheduled") c_function() ;
|
|||
|
(1.0 , "scheduled") my_ball.print_state() ;
|
|||
|
}
|
|||
|
|
|||
|
}
|
|||
|
```
|
|||
|
|
|||
|
|
|||
|
#### Job Calling Arguments
|
|||
|
|
|||
|
Job calling arguments adhere to C++ calling argument standards.
|
|||
|
|
|||
|
### Sim Object Methods
|
|||
|
|
|||
|
Methods may be defined within a sim object. These methods may be used as simulation jobs.
|
|||
|
A possible use for sim object methods would be to call non C/C++ code with minimal overhead from
|
|||
|
the S_define file.
|
|||
|
|
|||
|
### Specifying Scheduled Loop Job Class Order
|
|||
|
|
|||
|
This section of the S_define (encapsulated by "job_class_order{...};) can be used to specify a new
|
|||
|
scheduled loop job class order. The user may simply re-order the existing job classes that exist or
|
|||
|
can specify a new set of scheduled loop job classes. Job classes that are eligible for reassignment
|
|||
|
are listed in Table SD_1 between automatic and automatic_last inclusive. The order they are shown
|
|||
|
in the table is the default ordering.
|
|||
|
|
|||
|
```C++
|
|||
|
job_class_order {
|
|||
|
my_job_class_1 ;
|
|||
|
my_job_class_2 ;
|
|||
|
scheduled ;
|
|||
|
my_job_class_3 ;
|
|||
|
}
|
|||
|
```
|
|||
|
|
|||
|
### Simulation Object C++ properties
|
|||
|
|
|||
|
Sim objects are C++ objects. They possess all of the capabilities of C++ objects. This section
|
|||
|
describes how to use some C++ features with sim objects.
|
|||
|
|
|||
|
#### Simulation Object Instantiations
|
|||
|
|
|||
|
##### Multiple Instantiations
|
|||
|
|
|||
|
Sim objects are instantiated within the S_define file. They are regular class objects, and as such
|
|||
|
are treated that way. Sim objects may be multiply instantiated. Multiply instantiated sim objects
|
|||
|
works with both C and C++ models contained within the sim object.
|
|||
|
|
|||
|
```C++
|
|||
|
class ballSimObject : public Trick::SimObject {
|
|||
|
public:
|
|||
|
Ball obj ;
|
|||
|
ballSimObject() {
|
|||
|
("initialization") obj.state_init() ;
|
|||
|
("derivative") obj.force_field() ;
|
|||
|
("derivative") obj.state_deriv() ;
|
|||
|
("integration", &my_integ) trick_ret = obj.state_integ() ;
|
|||
|
(10.0, "scheduled") trick_ret = obj.state_print() ;
|
|||
|
}
|
|||
|
}
|
|||
|
|
|||
|
// Make 2 balls
|
|||
|
ballSimObject ball ;
|
|||
|
ballSimObject ball2 ;
|
|||
|
```
|
|||
|
|
|||
|
##### Sim Object Constructor Arguments and Initializer Lists
|
|||
|
|
|||
|
Sim objects instantiations may take arguments. These arguments may be used in the sim object's
|
|||
|
initialization list. An initialization list constructs member variables of the class. They
|
|||
|
are listed as a comma separated list after the declaration of a constructor. Arguments passed
|
|||
|
to the sim object constructor may be passed onto member variable constructors.
|
|||
|
|
|||
|
C structures may be zeroed out when included in the sim object's initialization list.
|
|||
|
|
|||
|
```C++
|
|||
|
class ballSimObject : public Trick::SimObject {
|
|||
|
public:
|
|||
|
Ball obj ;
|
|||
|
C_STRUCT c_struct ;
|
|||
|
|
|||
|
// passes int argument to obj constructor. Zeros out c_struct.
|
|||
|
ballSimObject(int num) : obj(num) , c_struct() {
|
|||
|
} ;
|
|||
|
}
|
|||
|
|
|||
|
// Sim object constructor requires an integer argument.
|
|||
|
ballSimObject ball(5) ;
|
|||
|
```
|
|||
|
|
|||
|
##### Sim Object Constructor Arguments and Job Control
|
|||
|
|
|||
|
Arguments to sim objects may also be used to control job execution. Most items in the job
|
|||
|
specification may be set to the value of an argument.
|
|||
|
|
|||
|
```C++
|
|||
|
class ballSimObject : public Trick::SimObject {
|
|||
|
public:
|
|||
|
Ball obj ;
|
|||
|
ballSimObject(int phase, double cycle , const char * job_class) {
|
|||
|
(job_class) obj.state_init() ;
|
|||
|
Pphase ("derivative") obj.force_field() ;
|
|||
|
("derivative") obj.state_deriv() ;
|
|||
|
("integration", &my_integ) trick_ret = obj.state_integ() ;
|
|||
|
(cycle, "scheduled") trick_ret = obj.state_print() ;
|
|||
|
}
|
|||
|
}
|
|||
|
|
|||
|
ballSimObject ball(1 , 10.0 , ~@~\initialization~@~]) ;
|
|||
|
// This ball has different job properties than the first ball.
|
|||
|
ballSimObject ball2( 2 , 5.0 , ~@~\default_data~@~] ) ;
|
|||
|
```
|
|||
|
|
|||
|
#### Multiple Constructors
|
|||
|
|
|||
|
Sim objects may define multiple constructors. Each constructor may define different job
|
|||
|
parameters or even an entirely different set of jobs. Arguments to the sim object
|
|||
|
instantiation determine which sim object constructor and which jobs are run.
|
|||
|
|
|||
|
```C++
|
|||
|
class ballSimObject : public Trick::SimObject {
|
|||
|
public:
|
|||
|
Ball obj ;
|
|||
|
ballSimObject(int phase, double cycle , const char * job_class) {
|
|||
|
(job_class) obj.state_init() ;
|
|||
|
Pphase ("derivative") obj.force_field() ;
|
|||
|
("derivative") obj.state_deriv() ;
|
|||
|
("integration", &my_integ) trick_ret = obj.state_integ() ;
|
|||
|
(cycle, "scheduled") trick_ret = obj.state_print() ;
|
|||
|
}
|
|||
|
ballSimObject(const char * job_class) {
|
|||
|
(job_class) obj.state_init() ;
|
|||
|
}
|
|||
|
|
|||
|
}
|
|||
|
|
|||
|
ballSimObject ball(1 , 10.0 , "initialization") ;
|
|||
|
ballSimObject ball2( "default_data" ) ;
|
|||
|
```
|
|||
|
|
|||
|
#### Sim Object Inheritance
|
|||
|
|
|||
|
Sim objects may inherit from other sim objects. Jobs in the derived class will be run after those
|
|||
|
of the base sim object class. Both C and C++ jobs may be inherited.
|
|||
|
|
|||
|
```C++
|
|||
|
class ballSimObject : public Trick::SimObject {
|
|||
|
public:
|
|||
|
Ball obj ;
|
|||
|
ballSimObject() {
|
|||
|
(10.0, "scheduled") trick_ret = obj.state_print() ;
|
|||
|
}
|
|||
|
}
|
|||
|
|
|||
|
class anotherBallSimObject : public ballSimObject {
|
|||
|
public:
|
|||
|
anotherBallSimObject() {
|
|||
|
// This job will run after the above state_print()
|
|||
|
(10.0, "scheduled") another_print() ;
|
|||
|
}
|
|||
|
}
|
|||
|
|
|||
|
anotherBallSimObject ball() ;
|
|||
|
```
|
|||
|
|
|||
|
#### Polymorphism in Sim Object jobs.
|
|||
|
|
|||
|
Polymorphism can be used to dynamically set objects at initialization or even change object types
|
|||
|
during runtime. Given an abstract class and two derived classes:
|
|||
|
|
|||
|
```C++
|
|||
|
class Ball {
|
|||
|
public:
|
|||
|
virtual void print_type() = 0 ;
|
|||
|
} ;
|
|||
|
|
|||
|
class Baseball : public Ball {
|
|||
|
public:
|
|||
|
virtual void print_type() ;
|
|||
|
} ;
|
|||
|
|
|||
|
class Soccerball : public Ball {
|
|||
|
public:
|
|||
|
virtual void print_type() ;
|
|||
|
} ;
|
|||
|
```
|
|||
|
|
|||
|
We may use a Ball base pointer in the S_define.
|
|||
|
|
|||
|
```C++
|
|||
|
class ballSimObject : public Trick::SimObject {
|
|||
|
public:
|
|||
|
Ball * ball_ptr ;
|
|||
|
ballSimObject() {
|
|||
|
(1.0 , "scheduled") ball_ptr->print_type() ;
|
|||
|
}
|
|||
|
} ;
|
|||
|
|
|||
|
ballSimObject ball ;
|
|||
|
```
|
|||
|
|
|||
|
`ball_ptr` is not instantiated when the simulation is compiled. If nothing is assigned to `ball_ptr`
|
|||
|
before the first scheduled call of `ball_ptr->print_type()` then the call will fail and the sim
|
|||
|
will core. We can allocate `ball_ptr` in the input file. We can even change `ball_ptr` in the
|
|||
|
middle of the simulation.
|
|||
|
|
|||
|
```python
|
|||
|
ball.ball_ptr = trick.TMM_declare_var_s("Soccerball[1]")
|
|||
|
trick.add_read(20.0 , """ball.ball_ptr = trick.TMM_declare_var_s("Baseball[1]")""")
|
|||
|
```
|
|||
|
|
|||
|
### State Integration Specification
|
|||
|
|
|||
|
Trick manages state integration with exceptional flexibility. The integration specification
|
|||
|
allows the developer to group the derivative, integration, and dynamic_event class modules
|
|||
|
(for any combination of sim objects) for state integration using a particular integrator and
|
|||
|
state integration time step. Some simulations will have several different sets of state
|
|||
|
parameters, each set requiring a unique state integration scheme and integration time step.
|
|||
|
Likewise, other simulations will require all the derivative class modules from a group of
|
|||
|
sim objects to be executed before the integration class modules of the same sim object group.
|
|||
|
The integration specification provides this capability.
|
|||
|
|
|||
|
The integration specification is of the following form:
|
|||
|
|
|||
|
```
|
|||
|
integrate <integrator_name> (<integration_dt>) <sim_object_name> [,<sim_object_name_n>] ;
|
|||
|
```
|
|||
|
|
|||
|
An alternative instantiation syntax which is pure C++ is of the form:
|
|||
|
|
|||
|
```C++
|
|||
|
IntegLoopSimObject <integrator_name> (<integration_dt>, <sim_object_name> [,<sim_object_name_n>], NULL ) ;
|
|||
|
```
|
|||
|
|
|||
|
This form must have NULL as the final argument to the instantiation.
|
|||
|
|
|||
|
The integrate tag is a reserved word for the CP. The <integration_dt> is a state integration
|
|||
|
cycle time in seconds. At least one sim object name must be specified followed by any number
|
|||
|
of additional sim object names separated by commas. An S_define can have at most one integrate
|
|||
|
statement per sim object, and at least one integrate statement for all sim objects.
|
|||
|
|
|||
|
### Parameter Collection Mechanism
|
|||
|
|
|||
|
The parameter collection mechanism is probably the most difficult capability of the CP to
|
|||
|
understand and utilize. This capability is useful when the user wants a single job to handle
|
|||
|
`n` number (either a known or unknown `n`) of parameters of the same type. The parameter
|
|||
|
collection mechanism is an alternative for a variable calling argument list The collection
|
|||
|
mechanism syntax in the S_define file is as follows:
|
|||
|
|
|||
|
```C++
|
|||
|
collect <reference> = { } ;
|
|||
|
```
|
|||
|
|
|||
|
or
|
|||
|
|
|||
|
```C++
|
|||
|
collect <reference> = { <reference_1> [,<reference_n>] } ;
|
|||
|
```
|
|||
|
|
|||
|
There is also a C code equivalent to adding collect references one at a time that may
|
|||
|
be put in a create_connections section of the S_define file. The advantage of this
|
|||
|
method is that not all of the collects must be listed in a single collect statement.
|
|||
|
This function call syntax may also be used in the input file to add collects at runtime.
|
|||
|
|
|||
|
```C++
|
|||
|
void create_connections() {
|
|||
|
reference = add_collect( reference , reference_1 ) ;
|
|||
|
reference = add_collect( reference , reference_2 ) ;
|
|||
|
}
|
|||
|
```
|
|||
|
|
|||
|
The collect capability allows the developer to build a job which accesses an unknown number
|
|||
|
of independent simulation parameters. For example, if designed correctly, a derivative class
|
|||
|
module should be capable of incorporating any number of known and unknown (future capabilities)
|
|||
|
external forces and torques without any code changes to the derivative module. The collection
|
|||
|
mechanism stores the addresses of, and number of, any number of independent parameters in a
|
|||
|
single pointer array. The derivative module can use this array to access each parameter in the
|
|||
|
collection list. See Section 10.0 for programming requirements for this capability.
|
|||
|
|
|||
|
The collect statements in the S_define file must be supported by source code implemented by
|
|||
|
the math model developer. This collect support code can reside in any function in the simulation,
|
|||
|
including functions that are not listed in the S_define file. In general, for every collect
|
|||
|
statement in the S_define file, there are two places where source code must be developed: a
|
|||
|
data structure definition file (`*.h`) and a function source file (`*.c`).
|
|||
|
|
|||
|
As a real world example, orbital dynamics can include a large number of environmental effects
|
|||
|
for high precision state propagation, or a small number of effects for general purpose state
|
|||
|
propagation. A spacecraft EOM derivative module should be designed to include any number and
|
|||
|
combination of known and unknown (future) effects. A low fidelity parameter collection for
|
|||
|
external torques on the spacecraft might look like:
|
|||
|
|
|||
|
```C++
|
|||
|
collect shuttle.orbital.rotation.external_torque[0] = {
|
|||
|
shuttle.aero.out.torque[0] } ;
|
|||
|
```
|
|||
|
|
|||
|
A higher fidelity parameter collection might look like:
|
|||
|
|
|||
|
```C++
|
|||
|
collect shuttle.orbital.rotation.external_torque[0] = {
|
|||
|
shuttle.aero.out.torque[0] ,
|
|||
|
shuttle.grav.out.gravity_gradient_torque[0] ,
|
|||
|
shuttle.solar_pressure.out.torque[0] } ;
|
|||
|
```
|
|||
|
|
|||
|
For those cases when there are no parameters to collect:
|
|||
|
|
|||
|
```C++
|
|||
|
collect shuttle.orbital.rotation.external_torque[0] = { } ;
|
|||
|
```
|
|||
|
|
|||
|
The key here is that if a new external torque for the spacecraft is added to the simulation,
|
|||
|
that torque can be accessed by the existing derivative module without code modification to the
|
|||
|
derivative module. Note that all parameters listed must be of the same type and array dimension.
|
|||
|
|
|||
|
To use the parameter collection mechanism of the S_define file, the developer must perform three tasks:
|
|||
|
|
|||
|
1. from the example above, the external_torque parameter must be declared in its data structure
|
|||
|
definition file as a two dimensional void pointer, i.e. `void ** external_torque ;`,
|
|||
|
|
|||
|
2. a loop must be placed in the derivative module which accesses the collected parameters, and
|
|||
|
|
|||
|
3. the parameter collection statement must be added to the S_define.
|
|||
|
|
|||
|
The external_torque parameter must be declared as a two dimensional void pointer for two reasons.
|
|||
|
First, the void type is not processed by the ICG. This means that this parameter cannot be recorded
|
|||
|
for output or assigned data for input. If the type was any other type than void, the ICG would
|
|||
|
assume the parameter required dynamic memory allocation and the resulting ICG generated code would
|
|||
|
cause a fatal runtime error (usually accompanied by a core dump). Second, from an automatic code
|
|||
|
generation viewpoint, the external_torque parameter is actually an unconstrained array of pointers,
|
|||
|
where the pointers in the unconstrained array could be of any type (including data structure pointers);
|
|||
|
i.e. the first pointer (*) of the declaration is the array dimension, the second is the address to
|
|||
|
each of the collected parameters.
|
|||
|
|
|||
|
To make the collection mechanism work, the developer must add specific collection mechanism code to
|
|||
|
their module. For the above example, the derivative module code might look like the following; the
|
|||
|
text in bold indicates code which will be unchanged regardless of the parameters being collected:
|
|||
|
|
|||
|
```C++
|
|||
|
#include "dynamics/v2/dynamics.h"
|
|||
|
#include "sim_services/include/collect_macros.h"
|
|||
|
|
|||
|
int derivative_job( DYN_ROT * R ) {
|
|||
|
|
|||
|
int i ;
|
|||
|
double **collect ;
|
|||
|
double total_torque[3] ;
|
|||
|
|
|||
|
total_torque[0] = total_torque[1] = total_torque[2] = 0.0 ;
|
|||
|
|
|||
|
/* typecast the void ** as a usable double** */
|
|||
|
collect = (double**)R->external_torque ;
|
|||
|
|
|||
|
/*
|
|||
|
Loop on the number of collected items
|
|||
|
from the above collect statement example:
|
|||
|
collect[0] -> shuttle.aero.out.torque
|
|||
|
collect[1] -> shuttle.grav.out.gravity_gradient_torque
|
|||
|
collect[2] -> shuttle.solar_pressure.out.torque
|
|||
|
*/
|
|||
|
for( i = 0 ; i < NUM_COLLECT(collect) ; i++ ) {
|
|||
|
total_torque[0] += collect[i][0] ;
|
|||
|
total_torque[1] += collect[i][1] ;
|
|||
|
total_torque[2] += collect[i][2] ;
|
|||
|
}
|
|||
|
|
|||
|
return( 0 ) ;
|
|||
|
}
|
|||
|
```
|
|||
|
|
|||
|
Several aspects of this example code which need mentioning are listed below.
|
|||
|
|
|||
|
1. A local pointer parameter must be declared of the same type as the parameters being
|
|||
|
collected, in this case the parameters being collected are double precision; hence,
|
|||
|
`double **collect ;`.
|
|||
|
|
|||
|
2. The `shuttle.orbital.rotation.external_torque` (actually a `void**`) is typecast as a
|
|||
|
`double**` and assigned to the local variable: `collect = (double**)R->external_torque ;`.
|
|||
|
|
|||
|
3. The number of parameters collected is saved in the first eight bytes before the
|
|||
|
address to the `external_torque` parameter. The conditional statement of the for loop
|
|||
|
demonstrates how the number of collected parameters is retrieved: `NUM_COLLECT(collect)`.
|
|||
|
4. This example, and all other collection mechanism code implementations, assume the
|
|||
|
developer knows the type and array size of the parameters being collected. In this
|
|||
|
example, the parameters collected were single dimensioned double precision arrays with
|
|||
|
three elements per array.
|
|||
|
|
|||
|
### Create Connections
|
|||
|
|
|||
|
The create_connections section contains arbitrary code that is executed right after sim
|
|||
|
object instantiations. Code in this section is performed before any job of any job class.
|
|||
|
The intended use of this section is to glue the sim objects together. Sim objects that
|
|||
|
need pointers to other sim objects may be assigned in the create_connections routine.
|
|||
|
Default parameters may also be set such as defining a default simulation stop time. Any
|
|||
|
arbitrary code may be added to the create_connections section.
|
|||
|
|
|||
|
There may be multiple create_connection sections in the S_define file. They will be
|
|||
|
concatenated together in the order they appear in the S_define file.
|
|||
|
|
|||
|
```C++
|
|||
|
class AsimObject : public Trick::SimObject {
|
|||
|
public:
|
|||
|
modelA a ;
|
|||
|
// This pointer points to a different sim object
|
|||
|
modelB * b_ptr ;
|
|||
|
|
|||
|
AsimObject() {
|
|||
|
// This job requires type modelB from a different sim object
|
|||
|
(1.0 , "scheduled") a.job(b_ptr) ;
|
|||
|
}
|
|||
|
} ;
|
|||
|
|
|||
|
class BsimObject: public Trick::SimObject {
|
|||
|
public:
|
|||
|
modelB b ;
|
|||
|
};
|
|||
|
|
|||
|
AsimObject A ;
|
|||
|
BsimObject B ;
|
|||
|
|
|||
|
void create_connections() {
|
|||
|
|
|||
|
// Connects the AsimObject and BsimObject together.
|
|||
|
A.b_ptr = &B.b
|
|||
|
|
|||
|
// Sets a default stop time for the simulation.
|
|||
|
exec_set_terminate_time(300.0) ;
|
|||
|
}
|
|||
|
```
|
|||
|
|
|||
|
[Continue to Making The Simulation](Making-the-Simulation)
|