Compare KleidiCV Gaussian blur performance across Neon, SVE2, and SME on Android
Introduction
Set up the Android build environment
Run the standalone SME Gaussian blur example
Build and customize the Gaussian blur performance explorer
Compare Neon, SVE2, and SME performance
Next Steps

  Log an issue

  Fork and edit

  Discuss on Discord

Understand the Gaussian blur example

examples/extract_one_operation/example_usage.c creates a 20x20, single-channel image containing a white vertical line on a black background. It calls sme_gaussian_blur_u8 with a 15x15 Gaussian kernel and KLEIDICV_BORDER_TYPE_REFLECT, then prints the output pixel values.

The companion CMake project builds sme_gaussian_blur from the SME Gaussian blur source and its small C API wrapper. It uses -march=armv9-a+sme, rather than linking the full KleidiCV library or using runtime dispatch.

Push and run the binary

Push the binary to a writable directory on the device:

    

        
        
adb push build/extract-android/example_usage /data/local/tmp/
adb shell chmod 755 /data/local/tmp/example_usage
adb shell /data/local/tmp/example_usage

Before the filter runs, every input row contains one white pixel, 255, at column 10 and zeros elsewhere. The 20 input rows are identical. The first three rows are as follows:

    

        
        0  0  0  0  0  0  0  0  0  0  255  0  0  0  0  0  0  0  0  0
0  0  0  0  0  0  0  0  0  0  255  0  0  0  0  0  0  0  0  0
0  0  0  0  0  0  0  0  0  0  255  0  0  0  0  0  0  0  0  0

The program prints 20 identical pixel rows. The first three rows of the output are similar to:

    

        
        Raw pixel values for the blurred output:
0  0  0  1  3  6  12  20  30  36  40  36  30  20  12  6  3  1  0  0
0  0  0  1  3  6  12  20  30  36  40  36  30  20  12  6  3  1  0  0
0  0  0  1  3  6  12  20  30  36  40  36  30  20  12  6  3  1  0  0

The original white line is at column 10. The 15x15 Gaussian kernel spreads it from columns 3 through 17, with the highest value, 40, remaining at the center. Every row is identical because the input is a vertical line and the REFLECT border mode preserves that pattern at the top and bottom edges.

If the device uses heterogeneous CPU clusters, pin the process to a known SME-capable CPU. For example, CPU 7 has the affinity mask 80:

    

        
        
adb shell 'taskset 80 /data/local/tmp/example_usage'

The hexadecimal mask is device-specific. Use the CPU topology of your own device when choosing an affinity mask.

What you’ve accomplished and what’s next

You’ve run the standalone SME Gaussian blur example and confirmed the filter output.

Next, you’ll build a performance explorer that controls the example’s kernel size and measurement count from the command line.

Back
Next