Code Generation for S-function - Matlab Interface

BilloBalloAldi · November 28, 2019, 2:58pm

Hi guys,

I am trying to use acados.

My goals are:

generate the code through acados Matlab Interfaces
run it from MATLAB

Afterwords I would like to port it to an embedded platform by using S-function.
This will in theory allow me to directly deploy the code onto my embedded platforms.
Therefore:
3) Generate the code for integration into a S-function
4) Run a Simulink model using the S-function

I need advices on where to start looking at.

For the points 1) and 2) I am able to run the examples in the folder
…/acados/examples/acados_matlab_octave/

For points 3) and 4) I don’t even know where to start

Thanks a lot,

BilloBallo

FreyJo · November 28, 2019, 4:01pm

Hi BilloBallo,

the S-function generation is happening through the template interface.
If you were able to set up an acados ocp solver ocp, you can run ocp.generate_c_code.
It will generate C code with the ocp formulation and options specified.
Moreover, it generates a main function, Makefile, Headers and an S-function.
To go embedded, you can also use the Code without the S-function, which might be easier depending on the plattform.
For the S-function, navigate into c_generated_code and run make_sfun to obtain the compiled S-function.
You can put this S-function into the S-function block of Simulink.
Note, that not all settings are supported by the interfaces yet, but feel free to post again, if you need something specific.
There is an ongoing Pull Request in which aims to improve the S-function generation, see https://github.com/acados/acados/pull/518.

Cheers,
Jonathan

BilloBalloAldi · November 29, 2019, 2:43pm

Hi Jonathan,
thanks a lot for your reply.
I was able to set up the acados ocp solver with matlab interface, but did not generate the code successfully.

I have added to the example file:
“ACADOS_ROOT_PATH/examples/acados_matlab_octave/linear_mass_spring_model/example_ocp.m”
ocp.generate_c_code() at the end of the script and gave me back this error:

#######################################################################
Error using ocp_generate_c_code (line 15)
mex templating does only support linear_ls cost for now. Got cost_type: ext_cost, cost_type_e: ext_cost.
Notice that it might still be possible to solve the OCP from MATLAB.

Error in acados_ocp/generate_c_code (line 144)
ocp_generate_c_code(obj)

Error in example_ocp (line 338)
ocp.generate_c_code()
#######################################################################

Since the error is saying “…still be possible to solve the OCP from MATLAB…” and not generate code I guess Matlab interface does not support code generation yet, am I right?

Do I need to use python only “acados_template” interface for this?

I have installed the python package from the folder
"ACADOS_ROOT_PATH/interfaces/acados_template/
running python3 install .

Then, I tried to run from the folder
“ACADOS_ROOT_PATH/examples/acados_python/test”
python3 main_test.py
but it gives me back the following error

#######################################################################
sh: 1: python: not found
Traceback (most recent call last):
File “main_test.py”, line 58, in
raise Exception(“acados status = {} on test {}. Exiting\n”.format(status, parameters))
Exception: acados status = 32512 on test (‘LS’, ‘ERK’, ‘PARTIAL_CONDENSING_HPIPM’, ‘SQP’). Exiting
#######################################################################

I have also added the env variable

export LD_LIBRARY_PATH=ACADOS_ROOT_PATH/lib

I hope I explained clearly what I did,
Thanks a lot.

BilloBallo

FreyJo · December 3, 2019, 4:34pm

Dear BilloBallo,

sorry I missed your reply for a while.

mex templating does only support linear_ls cost for now. Got cost_type: ext_cost, cost_type_e: ext_cost.
Notice that it might still be possible to solve the OCP from MATLAB.

this message is meant to say that it is supported by the MEX interface, but not by the templates.
This being sad, the templates do not (yet) support all possible options/ modules of acados, as the error says, it only supports linear least squares (linear_ls) as a cost module.
Here, you can find the problem description handled by acados, and the modules that can be used for cost, constraints and dynamics.
For the examples in <acados_root>/examples/acados_matlab_octave/getting_started, the code generation works, (at least on the GUC branch, which is to be merged soon), since they use more basic modules.
I hope you will give these another shot.

Just to note, the templates are the bottleneck for using more advanced acados modules.
I am not sure about the Python error.
But especially, if you want to use an S-function eventually, I would stick to Matlab.

Cheers,
Jonathan

FreyJo · December 4, 2019, 11:42am

Short addition to the linear mass spring example in Matlab, you can switch to using the modules supported by the templates easily.
Just use
cost_type = 'auto'
and
this constraint formulation

    % bounds on x and u
    	nbx = nx/2;
    	nbu = nu;
    	ng = 0;
    	nh = 0;

zanellia · December 27, 2019, 11:02am

Maybe an update just to avoid confusion: 1) you can use nonlinear least-squares when using the templates (simply they are not implemented in the MATLAB templating machinery - you can use NLSs from Python) 2) you can generate an S-Function from Python

Freddy · April 10, 2020, 2:53pm

Hello,
I really like the code generation for s-function option for Matlab.

Is there a way to output all calculated states from the ocp block, instead of just the first one like in the wind_turbine_nx6 example?
temp

FreyJo · April 15, 2020, 1:18pm

Hi @Freddy,

I am glad you like the S-function.

I think that the whole x trajectory in the OCP solver is not often needed, so I wont implement that.
However, please feel free to edit the template that is used for the S-function:

github.com

acados/acados/blob/d4eda65cc2671ca53adcc459c66fcd70702b806b/interfaces/acados_template/acados_template/c_templates_tera/acados_solver_sfun.in.c

/*
 * Copyright 2019 Gianluca Frison, Dimitris Kouzoupis, Robin Verschueren,
 * Andrea Zanelli, Niels van Duijkeren, Jonathan Frey, Tommaso Sartor,
 * Branimir Novoselnik, Rien Quirynen, Rezart Qelibari, Dang Doan,
 * Jonas Koenemann, Yutao Chen, Tobias Schöls, Jonas Schlagenhauf, Moritz Diehl
 *
 * This file is part of acados.
 *
 * The 2-Clause BSD License
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions are met:
 *
 * 1. Redistributions of source code must retain the above copyright notice,
 * this list of conditions and the following disclaimer.
 *
 * 2. Redistributions in binary form must reproduce the above copyright notice,
 * this list of conditions and the following disclaimer in the documentation
 * and/or other materials provided with the distribution.
 *

This file has been truncated. show original

In particular, you would have to edit these sections:

github.com

acados/acados/blob/d4eda65cc2671ca53adcc459c66fcd70702b806b/interfaces/acados_template/acados_template/c_templates_tera/acados_solver_sfun.in.c#L149




{%- if dims.nh > 0 %}
{%- set i_input = i_input + 1 %}
// lh
ssSetInputPortVectorDimension(S, {{ i_input }}, {{ dims.nh }});
{%- set i_input = i_input + 1 %}
// uh
ssSetInputPortVectorDimension(S, {{ i_input }}, {{ dims.nh }});
{%- endif %}


// specify dimension information for the output ports
ssSetOutputPortVectorDimension(S, 0, {{ dims.nu }} ); // optimal input
ssSetOutputPortVectorDimension(S, 1, 1 ); // solver status
ssSetOutputPortVectorDimension(S, 2, 1 ); // KKT residuals
ssSetOutputPortVectorDimension(S, 3, {{ dims.nx }} ); // first state
ssSetOutputPortVectorDimension(S, 4, 1); // computation times
ssSetOutputPortVectorDimension(S, 5, 1 ); // sqp iter


// specify the direct feedthrough status
// should be set to 1 for all inputs used in mdlOutputs
{%- for i in range(end=n_inputs) %}

github.com

acados/acados/blob/d4eda65cc2671ca53adcc459c66fcd70702b806b/interfaces/acados_template/acados_template/c_templates_tera/acados_solver_sfun.in.c#L353


    for (int ii = 0; ii < {{ dims.N }}; ii++)
        ocp_nlp_constraints_model_set(nlp_config, nlp_dims, nlp_in, ii, "uh", buffer);
{%- endif %}


    /* call solver */
    int rti_phase = 0;
    ocp_nlp_solver_opts_set(nlp_config, nlp_opts, "rti_phase", &rti_phase);
    int acados_status = acados_solve();




    /* set outputs */
    // assign pointers to output signals
    real_t *out_u0, *out_status, *out_sqp_iter, *out_KKT_res, *out_x1, *out_cpu_time;
    int tmp_int;


    out_u0          = ssGetOutputPortRealSignal(S, 0);
    out_status      = ssGetOutputPortRealSignal(S, 1);
    out_KKT_res     = ssGetOutputPortRealSignal(S, 2);
    out_x1          = ssGetOutputPortRealSignal(S, 3);
    out_cpu_time    = ssGetOutputPortRealSignal(S, 4);
    out_sqp_iter    = ssGetOutputPortRealSignal(S, 5);

Moreover, you will have to edit the Simulink block to adjust for the new outputs.

I hope that helps and isnt too involved…

Best,
Jonathan

DennisWeir · June 21, 2020, 12:50pm

Hi @FreyJo,

is this info still up to date? I searched the acados directory for the error message and couldn’t find it like that anymore, just for external costs.
My problem is that I want to control not a state but a performance output based on the states. (PO is a torque, states are torsions and rotary speeds of a drivetrain). Doing the calculation based on the states and writing that into y / y_e would be my goto based on what I see in the examples (haven’t tested it yet), but for the linear cost function I would need to neglect the nonlinear aspects in order to squeeze it into the linear cost term.
My goal is to simulate/test it in Simulink and finally run it on an ECU to see how it performs.

Cheers,
Dennis

FreyJo · June 22, 2020, 8:13am

Hi @DennisWeir,

I added the nonlinear least squares cost type for this part of the interface last week, following C Code doesnt generate for nonlinear_ls cost type

It sounds like this is just what you need, or do you also need external cost type?

Cheers,
Jonathan

DennisWeir · June 22, 2020, 9:15am

Hi Jonathan,

nonlinear least squares should work just fine, thank you very much!

Cheers,
Dennis