[Git-Donner]

Summary

Add an AArch64 NEON implementation of silk_LPC_analysis_filter, achieving 2.8–3.4× speed-up over the scalar C reference on Kunpeng 920.

Implementation

• Process 8 output samples in parallel using two NEON accumulators.
• Preserve the original 32-bit wraparound, rounding, and saturation behavior.
• Handle remaining samples with a scalar fallback.
• Integrate with the existing ARM NEON build infrastructure and retain the portable C reference.

Validation

• Bit-exact: 0 mismatches against the C reference across filter orders 8/10/12/16 and lengths 240/320/480, including random and near-saturation inputs.
• Build/tests: Fixed-point build, unit tests, and opus_demo round-trip completed successfully.
• Performance: 2.80–3.39× speed-up on Kunpeng 920 (GCC 12, -O3 -march=armv8-a). For d=16, len=480, execution time decreased from 2759 ns to 814 ns per call.

[Git-Donner]

Hi maintainers,

Just following up on this PR. All 6 GitHub Actions workflows are currently awaiting approval.

Could a maintainer approve the workflows when convenient? I'd also appreciate any feedback or review on the implementation.

Thanks :)

[jmvalin]

Is the PR AI-generated? If so, please let me know what you did and didn't do with the output.

[Git-Donner]

Yes, AI (Deepseek) was used to assist with this PR.

To clarify its involvement:

• What AI did: It helped explore the NEON optimization approach and generated an initial implementation draft for parts of the optimized code.
• What I did: I reviewed the generated code, modified and adapted it to match Opus's fixed-point arithmetic semantics, debugged the implementation, and handled the integration with the existing ARM build and dispatch infrastructure.
• Validation: I personally performed the bit-exact comparisons against the original C implementation, ran the build and unit tests, and conducted the performance benchmarks on ARM.

I did not submit the raw AI-generated output without review. I took responsibility for reviewing, modifying, debugging, and validating the final implementation.

The final code and the reported test results are my responsibility.

Thanks for bringing this up.