Example fbench

From wiki.emacinc.com
Revision as of 14:44, 11 December 2013 by Jgreene (talk | contribs)
Jump to: navigation, search
TODO: {{#todo:Review(12.03.13-14:18->JG+);(12.04.13-03:15->MD-);(12.04.13-13:56->JG+);(12.10.13-00:10->MD+)|Jgreene|oe 4,oe 5,jg,md,FinalDraft}}

This is a guide to the fbench C example project included in the EMAC OE SDK.

This example tests the speed and accuracy of a system's floating point operations. The project is a floating point benchmark and accuracy testing application that utilizes ray tracing algorithms and Fast Fourier Transforms to test your CPU and floating point library to its limits. It's also a good example of a method of processor performance comparison and compiler optimization testing. This project is an excerpt from the fbench project by John Walker of Fourmilab. See John Walker's Floating Point Benchmarks project homepage for more information.

The fbench project builds two executables: fbench and ffbench.

fbench is a trigonometry intensive floating point benchmark. It is a complete optical design raytracing algorithm, shorn of its ui.

ffbench is a Fast Fourier Transform benchmark. It loops through a fast Fourier transform of a square matrix of complex numbers, reverses the transform and then checks the results.

Opening, Building and Uploading the Project Files

1. Open the C/C++ editing perspective.

stub

2. Open the project files.

stub

3. Build the project.

stub

4. Upload the executables to the target system.

stub

Usage and Behavior

Hardware Requirements

The fbench project is intended for use on C implementations that define int as 32 bits or longer and permit allocation and direct addressing of arrays larger than one megabyte.

Using fbench

The fbench program is executed from the console. It takes a single optional parameter.

./fbench <itercount>

Where <itercount> specifies the number of iterations to be performed, with 1,000 being the default.
For archival purposes you'll want to use a value slightly higher than 1,000.

Usage Example

root@som9g20:/tmp# ./fbench 2000
Ready to begin John Walker's floating point accuracy
and performance benchmark.  2000 iterations will be made.


Measured run time in seconds should be divided by 2
to normalise for reporting results.  For archival results,
adjust iteration count so the benchmark runs about five minutes.

Press return to begin benchmark:

After fbench has finished it prompts the user to stop the timer (by pressing return).

Stop the timer:

Press return...

No errors in results.

...and fbench reports that no errors were found in the floating point operations.

A Note on Suspicious Systems

The default functionality as described above is for systems that can be trusted to be reliable. When working with a system that is suspected of having issues, fbench can be compiled with ACCURACY defined. This will generate a version that executes as an infinite loop, performs the ray trace and checks the results on every pass. All incorrect results will be reported. It will keep running until it is stopped manually (using, for instance, CTRL-C).

Using ffbench

The ffbench program is executed from the console. It takes no parameters.

./ffbench

Usage Example

root@som9g20:/tmp# ./ffbench
20 passes.  No errors in results.

It runs until it is finished and reports what it discovered. In this case it performed 20 passes (the default, specified in code) and found no errors.
The time that it takes for this benchmark to run is an indicator of the performance of the board running it. When running it from a Bash shell, the execution time can be measured thusly:

time ./fbench

Summary

The fbench floating point benchmark C example tests the speed and accuracy of your floating point operations, and is interactive by default. The ffbench example, on the other hand, is non-interactive by default and can be readily used both for benchmarking a board's floating point performance and to test the accuracy of its FPU.