Migrating x86_64 workloads to aarch64
Introduction
Porting methodology
Porting analysis
Development environment
Application porting
Run and evaluate
Evaluating performance on Arm hardware (Optional)
Review
Next Steps

  Log an issue

  Fork and edit

  Discuss on Discord

Run and evaluate

It is straight forward to compile and run the application, run the commands below:

    

        
        
            cmake -S src -B build
cd build/
make
./sobel_simd_opencv

A successful run will output execution time measurement results in microseconds in the terminal for the three implementations, and open four windows showing (as seen below) of the original, non-SIMD, SIMD and OpenCV images of a butterfly.

Image Alt Text:Sobel filter

In the results presented below, a value >1 is faster and a value <1 is slower in comparison to the normalized value. In short, higher values are better.

QEMU
CompilerGCC 11.3.0
Non-SIMD1.00
SIMD0.29
OpenCV0.02

The results in the table above have been normalized to the QEMU Non-SIMD value, giving the relative speed-up.

You observe the following:

  • the non-SIMD implementation is the fastest but does not reflect the performance on an aarch64 system
Note

Emulation does not give a representative view of how efficiently the algorithms run on Arm, it is only useful for functional purposes, not to measure performance

Closing notes

You have now ported an x86_64 application to aarch64, built and run the ported application on aarch64 using emulation, well done!

Continue to the next section to evaluate application performance on Arm hardware.

Back
Next