Matlab coder, simulink and concurrent execution

In this post we show how CasADi codegen can be integrated seemlessly with Matlab Coder, Simulink and parallel execution. This requires Casadi 3.8

Some context

We discussed Matlab Coder before, in the matlab Coder blog post. There, we hinted at the application of coder.target('MATLAB') to the Matlab System of the MPC blog post. Here, we implement that idea, and futher it to a robust thread-safe solution. The post does go a bit deeper into the details than usually.

Starting point

We use the Matlab System block developed in the earlier blogpost:

classdef casadi_block < matlab.System

    methods (Access = protected)
        function num = getNumInputsImpl(~)
            num = 2;
        end
        function num = getNumOutputsImpl(~)
            num = 1;
        end
        ...
        function setupImpl(obj,~,~)
            ...
        end

        function u = stepImpl(obj,x,t)
            
        end

        function resetImpl(obj)
            
        end
    end
end

Adding coder instructions

Next, we add coder instructions for incref and decref. These have been part of the CasADi codegen API since forever, yet they were always unused up to now, and may have been missing from previous examples or tutorials.

classdef fcn_system < matlab.System

    properties (Access = private, Nontunable)
        name
        src_name
    end

    methods (Access = protected)

        function num = getNumInputsImpl(obj)
                obj.name = 'foo';
                obj.src_name = 'bar';
                if coder.target('MATLAB')

                    % Codegen via a CasADi Function
                    F = opti.to_function(obj.name,{p},{x});

                    % Generate C code
                    F.generate([obj.src_name '.c'],struct('unroll_args',true,'with_header',true));
                
                    % Generate meta-data
                    config = struct;
                    config.sz_arg = F.sz_arg();
                    save([obj.src_name '_config.mat'],'-struct','config');
                end

            num = 1;
        end
        function num = getNumOutputsImpl(~)
            num = 1;
        end
        
        function setupImpl(obj,~,~)
            if coder.target('MATLAB')
               
            else
                coder.ceval([obj.name '_incref']);
            end

        end

        function releaseImpl(obj)
            if coder.target('MATLAB')

            else
                coder.ceval([obj.name '_decref']);
            end
        end
    end
end

We chose to piggyback our symbolic CasADi code and code-generation into getNumInputsImpl because that routine always gets evaluated in matlab interpreted context. This stands in contrast with setupImpl which may or may not get executed in an interpreted context depending on the Simulink settings.

We will eventually have multiple Simulink blocks running in parallel. Simulink makes sure that the setupImpl and releaseImpl happen in serial mode, which is required for incref/decref. This level of granularity seems to be beyond the capabilites of a Matlab function block, so that’s why we work with a Matlab System block.

A single-threaded system

We now have a Simulink model with two instances of the System. This could be a simple model for a cascaded/hierarchical MPC.

This will run fine with Normal, Accelerator or Rapid Accelerator modes.

Thread-safe checkout/release in codegen

This improvement can be found in the changelist.

This requires some changes in the code-gen:

F.generate([obj.src_name '.c'],struct('unroll_args',true,'with_header',true,'thread_safe',true));

And in the coder instructions:

coder.updateBuildInfo('addDefines','CASADI_MAX_NUM_THREADS=2');
coder.updateBuildInfo('addDefines','CASADI_THREAD_TYPE=CASADI_THREAD_TYPE_POSIX');

Note that we have only tested this with POSIX at the moment.

Multi-threading within a Simulink system

This section has nothing to do with CasADi really, but only with the “concurrent execution”

We follow the official cookbook to map each block to a separate thread.

Basically, we have one master Simulink model, model_task_par_top and then several sub-models each in their own slx file.

Validation

We use the following command to invoke Matlab coder, Simulink compilation and monitored running.

Simulink.architecture.profile('model_task_par_top');

The code generated by Matlab Coder should look familiar to advanced CasADi users:

Results

We gave a tour of advanced integration of Simulink with CasADi code-generation in a context of multi-threading.

Downloads: model_task.slx, fcn_system.m, model_task_par_top.slx, model_task_par_block1.slx, model_task_par_block2.slx, model_task_par_input.slx, model_task_par_output.slx