Optimize Opus decoding for ARM64

[zshang-oai]

Reduce decoder CPU time with portable specializations and an ARM64 NEON resampling kernel. On an Apple M4 Max, complete RFC stream replays with 24 kHz mono DecodeToInt16 output take 18.5–27.0% less time than the base revision.

Implementation

• Avoid integer divisions when decoding symbols from total-256 probability tables, with the generic path retained for other totals and range states.
• Specialize CELT deemphasis for mono 48 kHz and 24 kHz output while preserving filter updates for discarded samples.
• Select SILK excitation sign contexts once per 16-sample block and apply pseudorandom sign inversion with integer masking.
• Specialize 10/16-tap LPC synthesis and separate the first subframe's history-dependent warm-up from its contiguous tail.
• Partition LTP rewhitening and pitch prediction into contiguous history spans and boundary-crossing spans. Normalize coefficients once per subframe and use constant tap indices within contiguous spans.
• Keep up2HQ filter state in locals, simplify equivalent fixed-point multiplication, and add an eight-tap NEON FIR kernel. Assembly is selected by arm64 && go1.27 && !purego; other builds use the scalar implementation. The shared PCM quantizer is unchanged.

LTP correctness

Frame index zero separates saved history from the current frame. Rewhitening an output at i reads [i-dLPC, i], so only 0 <= i < dLPC needs mixed history access. Five-tap pitch prediction reads [i-lag-2, i-lag+2], so only lag-2 <= i < lag+2 needs mixed access: at most four output samples.

For a contiguous output span [A, B), the pitch helper receives source [A-lag-2, B-lag+2). Its local tap source[t+4-k] maps exactly to the original index A+t-lag+2-k. Rewhitening uses the corresponding dLPC preceding samples. The helpers preserve tap order, clamp-before-gain scaling, and ascending sample writes. Source and destination may overlap during pitch prediction, so each result is stored before the next output reads its history. The existing gain-ratio adjustment and frame-history handoff are preserved.

Measurements

Apple M4 Max, Go 1.27.0, 24 kHz mono DecodeToInt16; values are median microseconds per packet. Each implementation runs five rounds of 60 complete-stream replays, with rotated execution order, a fresh decoder per replay, and warm-up outside the timer. Reference libopus is the optimized Homebrew 1.6.1 build; its decoding loop runs in C without per-packet cgo calls.

Official stream	Base 1e54f79	Portable + NEON	libopus
12: mono, 20 ms	13.43	9.87	6.79
04: WB SILK	17.77	12.97	9.36
05: SWB hybrid	21.70	17.33	10.63
06: FB hybrid	24.15	19.69	12.08

Timings were collected before comment cleanup and equivalent loop rewrites for lint; the final cleanup has not been retimed. These are complete conformance-stream measurements on this host. The reference integer API uses soft clipping while Pion's integer API hard-clips. Ratios remain 1.39–1.63 times libopus; results do not establish performance on other ARM64 CPUs.

Validation

• Full repository TestRFC6716Conformance: 120/120 passed, covering all 12 streams, five sample rates, and mono/stereo. Latest ARM64 run after the lint fixes: Go 1.27.1, 73.189 s, no skips or failures, unchanged quality matrix.
• The full corpus also passed on the ARM64 purego path. NEON and portable output hashes matched for all 120 cases in the retained implementation.
• Linux/amd64 scalar fallback cross-build passed; git diff --check is clean.
• LPC unrolling can change compiler floating-point fusion. The cumulative change passes conformance but is not bit-identical to the base revision.

With the repository's RFC reference source and test vectors prepared:

OPUS_RFC6716_REFERENCE="$REFERENCE" \
OPUS_RFC6716_TESTVECTORS="$TESTVECTORS" \
go test -count=1 -timeout 20m -tags conformance \
  -run '^TestRFC6716Conformance$' -parallel 4 -v .

[codecov]

Codecov Report

❌ Patch coverage is 93.13725% with 21 lines in your changes missing coverage. Please review.
✅ Project coverage is 93.34%. Comparing base (1e54f79) to head (7d06453).

Files with missing lines	Patch %	Lines
internal/silk/lpc_synthesis.go	92.30%	7 Missing ⚠️
internal/silk/ltp_synthesis.go	89.28%	6 Missing ⚠️
internal/celt/synthesis.go	84.61%	3 Missing and 1 partial ⚠️
internal/silk/decoder.go	95.40%	3 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #239      +/-   ##
==========================================
- Coverage   93.37%   93.34%   -0.03%     
==========================================
  Files          58       61       +3     
  Lines       10710    10928     +218     
==========================================
+ Hits        10000    10201     +201     
- Misses        501      518      +17     
  Partials      209      209              

Flag	Coverage Δ
go	93.34% <93.13%> (-0.03%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[JoTurk]

last time I read ARM64 assembly or NEON SIMD I was still in uni years ago .. so I can't really review this I just hope it's correct and doesn't have bugs. I tested it yesterday and it seem able to compile and run.