PMU Metric Options

This table shows the options you can use to extract PMU metrics.

Check PMU Support on Target or Development Computer

The default driver provided for Linux-based target hardware uses the Performance Application Programming Interface (PAPI) for collecting performance counter information. This library provides an abstraction layer between the generated code and the kernel and PMU underneath. To check that you can use the PMU on your target or development computer:

Generate and Analyze Profile

The workflow for producing a profile using a PMU metric is similar to the workflow that produces an execution-time profile.

Configure the model for execution-time profiling.

openExample('ecoder/SILPILVerificationExample', ...
             supportingFile='SILTopModel.slx')

set_param(gcs, 'CodeExecutionProfiling', 'on');
set_param(gcs, 'CodeProfilingInstrumentation', 'coarse');
set_param(gcs,...
          'CodeProfilingSaveOptions', 'AllData');

set_param(gcs,...
          'CovEnable', 'off');
covSettings = get_param(gcs, 'CodeCoverageSettings');
covSettings.CoverageTool = 'None';
set_param(gcs, 'CodeCoverageSettings', covSettings);

Specify the metric that you want to extract from the PMU. For example, to extract L1 data cache misses, enter:

set_param(gcs, 'CodeProfilingCodeMetrics', 'l1dcm');

Run the simulation.

simOut = sim(gcs);

When the simulation is complete, in the Simulink^® Editor, the profiled model components are colored blue. For top-model SIL or PIL simulations, the Simulink Editor background is also colored blue. The simulation also generates the variable executionProfile (default) in the object simOut.

To view the results in the Code Profile Analyzer, in the Profiling details panel, click the Code Profile Analyzer link. Or, from the Command Window, run:

coder.profile.show(simOut.executionProfile)

The app displays information in the Function Execution and Summary panels.

In the Function Execution panel, the Function Metrics view provides a function-call tree and metrics for functions and their child functions.

The Function Metrics view provides this information:

The Relative Function Metrics view provides function metric values as percentages of caller function and total metric values. The table contains this information.

The Results section of the toolstrip provides features that you can use in your analysis of functions:

Compare Profile Against Baseline

You can compare metric values from the current simulation against values from a baseline simulation.

The panel displays:

The table cells are colored:

If the workspace variable for the baseline simulation does not contain data for a code section, the panel displays NaN.

If you click Compare in SDI, the Simulation Data Inspector uploads and compares the selected results.

To invert the comparison order, that is, make results from the current simulation the reference, click Invert order.

Register Custom Driver to Extract PMU Metrics

If your target hardware is not supported natively, you can register custom drivers to extract the required metrics from the PMU. Provided you define the API, you can extract metrics from the PMU of any target hardware. You are not limited to Linux-based target hardware or the PAPI library.

The Add profiling timer step in Use rtiostream API for PIL Target Connectivity shows how you can create a timer object for time retrieval. You can modify the step to enable the use of PMU metrics in code profiling.

Create a hardware counter object that describes the retrieval of counter information from the PMU. The object description must include a description of the function that retrieves counter values and its implementation.

This example uses a C function, pmu_counter, which returns a counter value as a uint64 data type.

HWCounterSignature = target.create('Function');
HWCounterSignature.Name = 'pmu_counter';
HWCounterSignature.ReturnType = 'uint64';

Capture the function in an API object.

HWCounterApi = target.create('API');
HWCounterApi.Functions = HWCounterSignature;
HWCounterApi.Language = target.Language.C;
HWCounterApi.Name = 'Hardware Counter API';

Capture the dependencies of the function, that is, the source and header files that are required to run the function.

HWCounterDependencies = target.create('BuildDependencies');
HWCounterDependencies.IncludeFiles = {'HWCounter.h'};
timerDependencies.IncludePaths = {'/home/user/Desktop/PMUDrivers'};
timerDependencies.SourceFiles = {'HWCounter.c'};

Create an object that combines the API and dependencies.

HWCounterImplementation = target.create('APIImplementation');
HWCounterImplementation.API = HWCounterApi;
HWCounterImplementation.BuildDependencies = HWCounterDependencies;
HWCounterImplementation.Name = 'Hardware Counter Implementation';

Create the hardware counter object and associate it with the L1 data cache miss counter.

HWCounter = target.create('HardwareCounter');
HWCounter.APIImplementation = HWCounterImplementation; 
HWCounter.Name = 'Hardware Counter';
HWCounter.Type = target.HardwareCounterType.L1DCMiss;

Append the new hardware counter to the timer and counters list in the processor object.

processor.Timers = [processor.Timers HWCounter];

The software supports these hardware counter types.