Compare Neon, SVE2, and SME performance

Run a comparison matrix

The following commands run 3000 measurements for 5x5, 7x7, 9x9, and 15x15 Gaussian blur on CPUs 0, 4, and 7:

    

        
        
for cpu_mask in 1 10 80; do
  for kernel in 5 7 9 15; do
    adb shell "taskset $cpu_mask /data/local/tmp/gaussian_blur_benchmark \
      --kernel $kernel --iterations 3000"
  done
done

    

The masks 1, 10, and 80 select the individual CPUs 0, 4, and 7 on the test device. Replace them with masks appropriate for your device.

Use p50 as the primary comparison metric. It’s the median of all measured calls, so it better represents typical per-call time than the mean when interrupts or frequency changes create occasional long calls.

Results from an SME-capable device

The table shows Neon p50 divided by SME p50. A value above 1.00x means SME is faster. Each result is from one process with a fixed CPU affinity and 3000 measured calls:

Understand the performance behavior

The 5x5 blur has inconsistent SME benefit. At 640x640, SME is slower on all three tested CPUs. The smaller workload doesn’t sufficiently amortize streaming-mode setup, loop overhead, border processing, and intermediate buffer management.

The 7x7, 9x9, and 15x15 kernels perform more work per pixel. This increases arithmetic intensity and lets the wider streaming-SVE vectors contribute more of the total execution time. The highest measured speedup is 2.87x for the 9x9 kernel at 1920x1080 on CPU 0.

The 15x15 results show an SME speedup for every tested CPU and resolution, ranging from 1.13x to 2.01x. Unlike the 3x3 through 9x9 fixed kernels, the 15x15 implementation doesn’t use a binomial variant. Its largest measured benefit is at 3840x2160 on CPU 7, where SME is 2.01x faster than Neon.

This doesn’t imply a fixed speedup for every device or image. The separable filter writes and reads an intermediate uint16_t buffer.

The following factors all affect the result:

• Cache capacity
• Memory bandwidth
• Streaming vector length
• Frequency scaling
• Thermal state

Don’t compare absolute times across CPU clusters. Compare Neon and SME within the same process, CPU affinity, kernel, and device state. For more confidence, run several independent processes and take the median of their p50 values.

What you’ve accomplished

You’ve built a standalone KleidiCV Gaussian blur example and measured explicit Neon, SVE2, and SME implementations. You’ve used CPU affinity to make comparisons more repeatable. You also identified that a larger kernel can expose more of the benefit of SME on an Arm-based Android device.