Skip to content

Viterbi optimization comparison

Curated source-specific evidence, not a new run or qualification of current source. Dates, measurements, failure outcomes and original source identifiers retain their recorded meaning. Old tags/assets have been retired and original commits may no longer resolve after the privacy rewrite. Where identifiers were sanitized earlier, the original hardware/run association cannot be independently recovered from the public derivative. Current acceptance lives in the roadmap.

Recorded and operator-accepted 2026-10-11 at d61cf03 / documentation ab4f5fa. Baseline code: b276cea518ddb9ab5f225653a9ce4f07432ba626; tracking head: f02bf79107bc9fdeeffa605e54a94e8145c4c1fa. The later CPU comparison uses this optimized baseline.

Decision and production scope

Recommendation: retain the narrowly scoped patch. Seven alternating warmed pairs show 20.41% ± 0.31 percentage points less complete native CPU, range 20.13–20.86%. Separately, actual browser memory64 workers show 12.37% ± 0.15 points less processing wall time, range 12.18–12.64%. No promotion claim is based on an isolated microbenchmark or CPU percentage. All promotion gates passed without material memory, block-spike, continuity or decoding regressions. These were agent qualification results; the operator subsequently accepted this completed patch and authorized the separate bounded follow-up.

Only HritReceiver::Impl::viterbi_decode changes: an immutable 128-byte encoder table for the existing polynomials and four exact integer costs per soft pair. All transitions/path metrics remain evaluated in their original order. Strict <, unreachable handling, final-state selection, survivors, traceback, warmup, hard disagreements, ASM/NRZ-M/randomization/RS, timing and bound evidence stay unchanged. No trellis restructuring, SIMD/NEON, FIR, threads, queues, dependencies, SDK, runtime policy or deployed legacy-browser changes. Phase 4 is unauthorized.

Frozen builds and environment

Baseline was archived before production edits, rebuilt separately, and reproduced the installed decoder hash exactly. Signed candidate code commit: d61cf03de9e0a528e0b606cfa40d81a8306281e3 (OpenPGP verification G). Current receiver source hash: e31d4ed29027ba075a76d3424efac1a3add524cff361b81edd3099aaceba0c57. Measured candidate binary remained identical after adding test-only trace detail.

Pi 5 Cortex-A76/aarch64, four cores, 7.9 GiB RAM, Debian Trixie, kernel 6.18.50+rpt-rpi-2712; g++ 14.2.0-19. Both native builds/harnesses use -O2 -std=c++20 -Wall -Wextra -pthread, no fast-math/PGO/architecture tuning. Only the benchmark process was pinned to CPU 3. Live baseline decoding continued. No global CPU policy change. Isolated SDK remains 4.0.23, legacy 3.1.69.

Artifact SHA-256
Frozen scalar receiver source 8d6ed9e4694216724830dcdbba4f1a29d1b8a1a4715490649f329eef4b6a109d
Installed/rebuilt baseline native decoder 6d7f03284c08ab6e547b51c5b48d7551c0775b675feff307e67129b01ceef886
Candidate native decoder 9541bc612f847031ec42d1ea55671335ca10aa43f7f1647a317663e77ea01ff1
Baseline benchmark executable 2e9f51940f0a5b8fb66a35c9826d50a0ea92079e4a6d582dd1a8e31b8f7c03c2
Candidate benchmark executable 9d4c861ceb9ce282a7909161d6e8b2623fd99fc4d780a8dae227d37cd0607b1f
Baseline memory64 receiver 1952710d63027eba09f52244c2ced0b883727154202b65dd7c2882b3c7b4737e
Candidate memory64 receiver e417e5acfaf22225e3da54151a4a650f6e1654a45a18009e227deab7e9c2c2d9
Separate memory64 differential artifact 6121662349f823d3c8d6b11bd9483641af1653b9f1de1167aec6d810f470523d

Uninstrumented new-path WASM 32,265 → 32,353 bytes, i64 memory/pointer ABI and exact hash/import/export allowlists inspected. These are qualification artifacts, not a deployment to the legacy browser path. Native nm identifies 128 bytes of read-only encoder data, no audit/scalar-reference symbols. Survivors remain 8,224×64 bytes; costs use a 16-byte local array, not a new heap allocation.

Executed correctness evidence

Frozen scalar reference retains independent parity/branch arithmetic and the original loop. It never uses the candidate table/helper. Instrumentation exists only in separate native, sanitized and receive-audit.wasm targets, not the production hot path.

  • All 128 mappings, 65,536 soft pairs × four costs exhaustively agree.
  • Every step's 64 metrics, reachable/unreachable transition, strict-improvement write/tie decision, written/untouched survivor, final state, pre-ASM/RS decoded bytes and hard disagreements agree.
  • Nine saturated/neutral/alternating/noisy soft vectors explicitly fail ASM; clean/corrected/uncorrectable/offset IQ supplements include a real RS rejection. Rejected candidates are compared before RS can conceal Viterbi differences.
  • Actual local ARM64 Chromium 154.0.8037.92 ran 113,383,822 differential checks, 105 candidates, 863,520 steps, 1,619,137 scalar ties, 262,272 arithmetic checks and 847 audit-worker checks (845 uninstrumented). Seccomp 2/NoNewPrivs 1; no CDP, experimental flags, sandbox bypass or wasm32.
  • Independent IQ preserves 92 ordered frames / 17 corrected symbols, hash 70e147554060f5be9aa117737a6061fcc3f54f46accd58b3f9608862e8450677. Native/browser compare every ordered byte to unchanged legacy decoding. Every lane agrees with the frozen native oracle: nonzero corrections only at zero-based frames 10 [1,0,0,1], 36 [0,0,0,1], 55 [4,4,2,4].
  • Baseline/candidate native manifests match bytes, per-lane corrections, hard disagreements and completion cursors. Audited 65,536-/1,022-byte chunk and sanitizer manifests also match exactly; instance/reset/discontinuity/bounds, clean/corrected/failed/offset cases and ASan/UBSan/leak checks passed.
  • Shared contract 4,083, browser glue 127 and tooling 11 checks passed. Gateway protected-queue regressions and mock HTTP/IPv4/TLS/IPv6/auth/concurrency/handoff integration passed. Hosted results are separately recorded below.

Independent source: Lucas Teske's GOES sample index, GOES-16 2017-01-17 16:25:08 UTC / 5 MS/s. The unchanged cached 33,554,432-byte prefix hash is b34eadddd172d8713be686d29b28541fe86fa3158bdf17f9b57a4ba48530a326. No full download/new live recording. Raw/generated payloads stay ignored.

Native complete-pipeline measurement

Harness includes the unchanged native service source and invokes its actual Decoder::process/publish/snapshot; the service entrypoint is never invoked. No SDR/socket/listener or acquisition authority. Both variants use identical harnesses/settings, chunks and output treatment.

Existing fixture conversion: signed 16-bit IQ → CU8 high byte; 16,777,194 bytes / 8,388,597 complex samples / 1.6777194 source seconds. A test-only 5-MS/s fixture adapter exercises the existing channelizer without relaxing installed job validation or changing the live 2.4-MS/s configuration.

Timing includes copying/validation, residual/NCO/FIR channelization, full demod/sync/Viterbi/NRZ-M/RS, evidence/hash, canonical record encode/decode/zero/ disposal, counters and serialized snapshots (one per fixture's first source-time second). Offline replay does not exercise kernel IPC/network; live checks do. Load/conversion/construction/filter setup/compilation/warmup are excluded.

Warm twice, then seven alternating pairs, 12 replays per measured run: 1,104 frames, 1,116 verified/disposed records, 1,763,628 wire bytes, 12 snapshots, 100,663,164 source samples. All fixed diagnostic objects agree, including carrier/ timing/power/lane totals and 1,298 hard disagreements per replay. No output failures/drops. Means ± sample SD across seven runs:

Run UUIDs and actual decode-UTC times legitimately differ; agreement concerns stable diagnostics, ordered payload/evidence/cursors and record counts/lengths, not an artificially identical runtime envelope hash.

Metric Baseline Candidate
CPU seconds / source second 0.75639 ± 0.00209 0.60197 ± 0.00097
CPU/source range 0.75399–0.75922 0.60056–0.60343
Processing speed / real time 1.32205 ± 0.00358× 1.66132 ± 0.00270×
Source throughput 6.610 MS/s 8.307 MS/s
User cycles / source sample 362.211 ± 0.983 288.215 ± 0.431
User instructions / source sample 913.859 ± 0.001 691.536 ± 0.001
Block p50 2.43945 ± 0.00309 ms 2.43954 ± 0.00481 ms
Block p95 9.52586 ± 0.04732 ms 6.69504 ± 0.01133 ms
Block p99 9.65569 ± 0.04618 ms 6.83112 ± 0.03229 ms
Maximum block, per-run mean 10.57496 ± 1.00462 ms 7.22704 ± 0.33394 ms
Worst observed block 12.65287 ms 7.97754 ms
Peak RSS, range 21,552–21,600 KiB 21,552–21,584 KiB

Native throughput improves 25.66%. The median stays essentially unchanged: non-frame blocks still do identical DSP. Expensive frame blocks improve. Observed gain is well beyond noise, with no material memory or spike regression. Temperatures approximately 64–69 °C, clocks 2.400 GHz, throttled 0x0; PMU time-enabled equals time-running, no multiplexing. PMU excludes kernel; thread CPU includes system time. Context switches/faults/load/startup/sensors are retained in build/viterbi-evidence/native-benchmark.json. RSS includes test input. Allocation call counts were not instrumented; no new Viterbi heap allocation is established by inspection. Offline IPC lag/queue is not manufactured as zero. File loading took 60.19–82.14 ms per process, excluded from processing; constructor setup summed 10.36–10.90 ms across 12 instances, also excluded. Whole-process resource accounting reports 2,351–2,355 minor / zero major faults; these include setup/input pages, not a measured steady-state allocation rate. Involuntary context switches were 168–180 baseline / 134–139 candidate. Competing live decoding remained roughly 0.75 cores in both recorded-input series; load and thermal/clock readings were sampled before and after each run.

Paired results (native CPU/source-second; browser gain is independently measured worker wall time). All seven native and seven browser pairs improved:

Pair / execution order Native baseline Native candidate Native CPU reduction Browser wall reduction
1 · baseline → candidate 0.75399 0.60217 20.14% 12.48%
2 · candidate → baseline 0.75627 0.60202 20.40% 12.64%
3 · baseline → candidate 0.75706 0.60343 20.29% 12.39%
4 · candidate → baseline 0.75507 0.60228 20.23% 12.38%
5 · baseline → candidate 0.75879 0.60056 20.85% 12.27%
6 · candidate → baseline 0.75429 0.60243 20.13% 12.27%
7 · baseline → candidate 0.75922 0.60088 20.86% 12.18%

No isolated Viterbi timing is used as the acceptance benchmark.

Actual browser memory64 results

Seven alternating warmed pairs, two replays/run, same local ARM64 Chromium, data/chunks and hash/ABI-verified artifacts. Ordered bytes, all lane corrections, 32 exact metric selectors and zero bounded output drops checked every run. Data/compilation/setup excluded. This measures channelizer+receive/FEC+bounded JS/WASM copies/drain, not native records/IPC. performance.now() reports worker wall time; per-worker CPU/PMU and browser-process peak RSS are unavailable.

Metric Baseline Candidate
Wall seconds / source second 1.03765 ± 0.00374 0.90925 ± 0.00196
Processing speed / real time 0.96373 ± 0.00346× 1.09982 ± 0.00237×
Block p50, mean 3.9893 ms 3.9807 ms
Block p95, mean 11.8821 ms 9.5764 ms
Block p99, mean ± SD 11.9607 ± 0.0431 ms 9.6729 ± 0.0420 ms
Worst observed block 22.370 ms 9.970 ms
Linear memory, every run 16 MiB 16 MiB

12.37% paired wall-time reduction, not a native-derived CPU estimate. This qualifies this Pi/browser/workload, not all browsers or multi-stream capacity. Actual >4-GiB addressing is separately pending Phase 12.

Before/after profiles and remaining hotspots

Payload-free 199-Hz user-cycle IP-only sampling, 30 replays each: baseline 7,596 samples, candidate 6,070, zero lost samples. No stack/register/ user-buffer capture. Sampling includes startup/load (<1% of these longer runs), unlike the processing stopwatch. Normalize each symbol's summed sample periods by 251,657,910 source samples. Self-cycle estimates, not substage timers:

Hotspot Before share After share Approx. cycles/source sample, before → after
Viterbi/survivors/traceback/hard check 48.14% 36.42% 174.54 → 105.57
Demod/AGC/carrier/SRRC 23.18% 28.68% 84.04 → 83.12
Channelizer/residual/NCO/FIR 18.25% 21.20% 66.18 → 61.44
Symbol/timing recovery 5.16% 7.16% 18.72 → 20.76
Four RS lanes 1.80% 2.16% 6.54 → 6.25
Input/mean checks + dispatch 2.11% 2.76% 7.65 → 8.00
PicoSHA block samples 0.33% 0.25% 1.19 → 0.72

Viterbi normalized sampled cost falls about 39.5%. Demod's larger share does not mean it slowed: absolute cost is essentially flat. Unchanged channelizer/ timing differences are sampling variability, not claimed improvements/regressions. Inlining/skid and small buckets limit attribution; the PicoSHA bucket alone is not the price of all records/copies/diagnostics. Raw profiles/symbol reports stay ignored under build/viterbi-evidence/{baseline,candidate}-profile.{data,txt}.

Ranked follow-ups — proposals ONLY

Benefits are untested estimates, not measured savings, and are not additive. Every item requires separate approval; none was implemented.

  1. Remaining Viterbi transitions: 36.42%, ~105.57 cycles/source sample. viterbi_decode still visits 128 transitions and copies 64 metrics/step. Paired-predecessor/butterfly organization might save 20–35% of this stage, 7–13% whole pipeline. Integer/tie/unreachable/traceback order risks are high. Require frozen per-transition reference, exhaustive/rejected/independent evidence, native/memory64/sanitizers and alternating full-pipeline benchmarks.
  2. Channelizer FIR/locality: 21.20%, ~61.44 cycles/source sample for the whole component, not FIR alone. hrit_channelizer.cpp::filter_at/process uses serial tap sums/ring accesses. Aligned/contiguous products or native vectors might yield 2–7% whole pipeline. Reduction/FMA changes can alter soft symbols/ frame acceptance; require summation-policy review, impulse/passband/DC/offset/ timing and broader noisy IQ, native/memory64 byte/evidence qualification.
  3. Demod/SRRC organization: 28.68%, ~83.12 cycles/source sample, including internal DSP AGC/carrier/RRC; hardware AGCs stay off. process_sample already pairs symmetric taps. Profile instruction locations before arithmetic-preserving I/Q/coefficient locality work; 1–5% whole pipeline speculative. Do not remove timing hypotheses or change inverse-sqrt/recovery precision. Require per-sample/ soft-symbol and downstream integer differential tests and independent noisy IQ.
  4. Survivor packing: 526,336 bytes/instance inside the Viterbi bucket; isolated CPU share unmeasured. Compact bits may improve cache locality but add read/ modify/write work. 0–3%, possibly negative estimated benefit. Portability, bit order, bounds, unwritten states and lifetime require sanitizer/trace tests.
  5. Records/hash/copies/diagnostics/IPC: small sampled buckets (PicoSHA 0.25%, not all record work). Measure allocations/call counts if future delivery makes this material. Bounded reusable scratch might save under 1–2%, not proven. Memory protection, ownership/cancellation and hash coverage risks outweigh current evidence for a rewrite. Keep encode/verify/disposal and validation.

Reproducible commands and evidence

From repository root; freeze the baseline before candidate builds:

mkdir -p components/rfsentry-core/build/viterbi-baseline-source
git archive b276cea518ddb9ab5f225653a9ce4f07432ba626 \
  components/rfsentry-core components/rfsentry-decoder components/rfsentry-api \
  | tar -x -C components/rfsentry-core/build/viterbi-baseline-source
RFSENTRY_GOES_OTA_WAV=/path/to/hash-verified-prefix.wav \
  components/rfsentry-core/tools/test.sh
./scripts/node-toolchain.sh run node \
  components/rfsentry-core/tools/viterbi-benchmark.mjs build
./scripts/node-toolchain.sh run node \
  components/rfsentry-core/tools/viterbi-benchmark.mjs measure
./scripts/node-toolchain.sh run node \
  components/rfsentry-core/tools/viterbi-benchmark.mjs profile
./scripts/node-toolchain.sh run node \
  components/rfsentry-core/tools/local-browser.mjs receive-benchmark

Browser measurement needs the frozen module/descriptor at build/viterbi-evidence/baseline-receive.wasm and build/viterbi-evidence/baseline-receive-artifact.json; saved before editing. To reproduce later, run archived baseline tools/build.mjs receive with CORE_EMSDK pointing to the existing pinned SDK, then copy its verified artifacts; never relabel candidate bytes as baseline. Measurement defaults seven pairs, 12 replays; bounded VITERBI_PAIRS/VITERBI_REPEATS optional. PMU/profile uses local sudo; browser needs neither sudo nor CDP. No global policy changes.

Limits

These are the recorded native/browser comparisons, not current-head results. Browser wall time is not CPU/PMU or peak RSS; profile attribution and queue/lag maxima are sampled. Live peaks are run-cumulative. >4-GiB execution, wider browsers and aggregate multi-stream capacity were not qualified. Deployment restart loss is separate from steady-state continuity; later delivery features do not retroactively change this experiment.