Repository navigation
perf: howl_detector per-stage bench; the guard's cost on Linux GCC - #87
Merged
Merged
Conversation
The guard costs ~4x on the Linux GCC runner what it costs on the Intel Mac (HowlGuardHost.CostPerBlock's comment, from PR #80's CI) while the canceller runs ~2x faster there. To localize that in one CI run the detector bench now times one block's stages separately, float and double, through the library's own code: howl_bank resonators + envelopes (a detector whose tick never comes, so process_block runs run_bank alone) howl_tick howl_detail::decide on envelope snapshots howl_tick_logs decide()'s 66 std::log10 calls alone howl_fit one band's least-squares fit howl_readouts harmonic / subharmonic readouts + 32 band_level_db guard policy howl_guard update() + apply(), the detector not run guard m1 / m2 the whole guard per block, as CostPerBlock times it The existing b64_f64 / b64_f32 cases are unchanged. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DDhgJqxUsdmWPKh6pKrV5A
Diagnosis only, reverted before the PR is ready: replicas of howl_detector::run_bank (verbatim, __restrict, an indexed select, resonators alone, envelopes alone, bands-outer), the howl stages at 5 x 0.3 s in bench-smoke (GCC), GCC's -fopt-info-vec report and disassembly of run_bank, and the same stages on Linux Clang and macOS. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DDhgJqxUsdmWPKh6pKrV5A
…ity harness Round 1 on the Linux GCC runner: run_bank is scalar there with a comiss/jbe branch on the envelope's attack/release select (10.2 us per block, float and double alike); __restrict on LOCAL pointers changed nothing, an indexed select took it to 4.3 us. Round 2 tries __restrict PARAMETERS, both arms computed before the select, and the coefficient by bit mask; dumps GCC's vectorizer notes for the replicas; and adds a bit-for-bit identity dump of every detector readout (main's header vs the branch's) for the fix to come. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DDhgJqxUsdmWPKh6pKrV5A
…orizes it The guard cost ~4x on the Linux GCC runner what it costs on the Intel Mac. The per-stage bench put all of it in run_bank, the 32 resonators and their envelopes: GCC 13.3 left the band loop scalar, with a comiss/jbe branch on the envelope's y > env pick that noise mispredicts, while Clang vectorized it (10103 vs 1060 ns per block, float, Linux GCC vs Linux Clang; double 10383 vs 2039). The tick (decide()) costs the same order on every host (730 ns float on GCC, 546 on Clang). On GCC, __restrict on the loop's LOCAL pointers changed nothing (10603 ns); an indexed, branch-free pick took the scalar loop to 4364 ns; the same loop with its arrays as __restrict PARAMETERS vectorized (1164 ns float, 2273 double), with the original ternary. So the loop moves, unchanged, into howl_detail::run_bank with restrict parameters; the member forwards to it. Each band's arithmetic is the old loop's, operation for operation. A dump of every readout after every process_block call (4 configs, 6 signals that drive every trigger path, irregular partitions; 3440844 values) is bit-identical between main's header and this one on AppleClang x86_64 (local); the PR runs the same dump on Linux GCC, Linux Clang and macOS CI. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DDhgJqxUsdmWPKh6pKrV5A
… macOS legs Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DDhgJqxUsdmWPKh6pKrV5A
The run_bank replicas, the identity dump and the TEMP CI steps (GCC vectorizer report, howl stages on Linux Clang and macOS, the identity comparison, the CostPerBlock print) did their job; their measurements are in bench/README.md and #87. ci.yml is main's again and bench_howl_detector.cpp is the per-stage bench of the first commit. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DDhgJqxUsdmWPKh6pKrV5A
bench/README.md records the per-stage breakdown on the Linux GCC runner before and after the bank's __restrict parameters (the bank 10291 -> 1288 ns float, 10182 -> 2434 double; the guard per block 10919 -> 2059 / 11250 -> 3407 at one mic), the cause, and the bit-identity check. howl_detector.h's COST paragraph and HowlGuardHost.CostPerBlock's measured comment gain the Linux GCC figures (CostPerBlock on job 113795845819: float 1998 / 3986 ns, double 3371 / 6793 ns at 1 / 2 mics). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DDhgJqxUsdmWPKh6pKrV5A
tap
marked this pull request as ready for review
October 9, 2026 14:28
tap
added a commit
that referenced
this pull request
Oct 9, 2026
The run_bank replicas, the identity dump and the TEMP CI steps (GCC vectorizer report, howl stages on Linux Clang and macOS, the identity comparison, the CostPerBlock print) did their job; their measurements are in bench/README.md and #87. ci.yml is main's again and bench_howl_detector.cpp is the per-stage bench of the first commit. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DDhgJqxUsdmWPKh6pKrV5A
tap
added a commit
that referenced
this pull request
Oct 9, 2026
bench_aec.cpp's -Wsign-conversion under -DMUTAP_WERROR=ON on AppleClang (CI's bench-smoke builds without -Werror), and run_bank's bit-identity not compared under the Cortex-M / Hexagon toolchains. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DDhgJqxUsdmWPKh6pKrV5A
tap
added a commit
that referenced
this pull request
Oct 9, 2026
bench_aec.cpp's -Wsign-conversion under -DMUTAP_WERROR=ON on AppleClang (CI's bench-smoke builds without -Werror), and run_bank's bit-identity not compared under the Cortex-M / Hexagon toolchains. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DDhgJqxUsdmWPKh6pKrV5A
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
HowlGuardHost.CostPerBlockrecorded the guard at 8575 / 12794 ns per block (float, 1 / 2 mics) on the Linux GCC runner (job 110465810942) against 2180 / 4983 ns on the Intel Mac, while the canceller ran ~2x faster there. This PR finds and removes the excess.What changed
bench/bench_howl_detector.cpp: a per-stage breakdown, float and double, through the library's own code:howl_bank(resonators + envelopes: a detector whose tick never comes),howl_tick(howl_detail::decideon envelope snapshots),howl_tick_logs(its 66std::log10calls),howl_fit,howl_readouts,bench_guard/policy_m1(the guard's own step, no detector) andbench_guard/m1,m2(the whole guard, asCostPerBlocktimes it). The existingb64_f64/b64_f32cases are unchanged.include/mutap/howl_detector.h: the bank loop moves, unchanged, intohowl_detail::run_bankwith__restrictparameters; the member forwards to it.bench/README.md(the table below and the cause), the detector's COST paragraph, theCostPerBlockmeasured comment.Temporary diagnostic commits (run_bank replicas, GCC's vectorizer report, the stages on Linux Clang and macOS runners, an identity dump, a
CostPerBlockprint) were reverted in3a0610c; their numbers are below.The breakdown (Benchmark smoke job, GCC 13.3, medians of 5 x 0.3 s, ns per block, float / double)
howl_bankhowl_tickhowl_tick_logshowl_fithowl_readoutsbench_guard/policy_m1howl_detector/b64bench_guard/m1bench_guard/m2Other hosts (TEMP job, same stages; before and after are different runner instances, so compare within a column only): Linux Clang before: bank 1073 / 2061, detector 1615 / 2746, guard m1 1633 / 2780; after: 1320 / 2445, 2000 / 3363, 2061 / 3432. macOS arm64 CI before: bank 1271 / 1217, detector 1516 / 1833, guard m1 1379 / 1524; after: 1203 / 1368, 1734 / 1708, 1623 / 1694.
HowlGuardHost.CostPerBlockafter (TEMP print, run 37922694658): Linux GCC float 1998 / 3986 ns (0.62 / 1.24 % of that runner's 321 us canceller), double 3371 / 6793 ns (1.03 / 2.07 %); Linux Clang float 2002 / 3990, double 3296 / 6561. The Intel Mac's recorded figure is 2180 / 4983 float.The cause
All of the excess was the bank. GCC 13.3 left the band loop scalar (
-fopt-info-vec: "couldn't vectorize loop", "not vectorized: no vectype for stmt" on the coefficient loads) and compiled the envelope's attack / release pick ascomiss/jbe, a branch white noise mispredicts; Clang vectorized the loop. Replicas of the loop on the GCC runner (float / double): verbatim 10103 / 10383;__restricton the local pointers 10603 / 10379 (no change); an indexed, branch-free pick 4364 / 4380 (scalar); the arrays as__restrictparameters 1164 / 2273 with the original ternary, the same with a bit-mask pick 1182 / 2311; the resonators alone 2545 / 2689, with restrict parameters 653 / 1283. GCC honours restrict on parameters, not on these locals.Proof of identical output
The loop's arithmetic is unchanged, operation for operation. A dump of every readout (tripped, verdict, trigger, confidence, peak / line Hz, prominence, growth per pass and per s, rise, harmonic and subharmonic ratios, level, power, every band level) after every
process_blockcall, over 4 configs x 6 signals that drive every trigger path (none / growth / ceiling / level = 65614 / 6824 / 19446 / 456 ticks) with irregular partitions, 3,440,844 values, compared main's header against this one withcmp: bit-identical on Linux GCC 13.3, Linux Clang and macOS arm64 (CI, run 37922694658) and AppleClang 17 x86_64 (local). Nothing the tests print moves: the 186 detector / guard / chain / two-mic / reverb tests (--gtest_filter='*Howl*:*howl*:*Afc*:*afc*:*TwoMic*:*two_mic*:*Reverb*:*reverb*') run on main's header and the fullcteston this branch (Intel Mac, AppleClang 17, Release,-DMUTAP_WERROR=ON; 420 / 420 passed) print the same 288 lines in the 74 tests that print anything, onceCostPerBlock's two wall-clock lines are set aside. The detector and the guard are in neither the fingerprint harness nor the icount scenarios; the ratchet and the fingerprint legs pass unchanged.On the Intel Mac (AppleClang already vectorized the loop) nothing moves beyond noise: four interleaved rounds at load ~7,
howl_bankfloat main 1395-1467 ns vs this branch 1410-1476 ns, the detector 2133-2330 vs 2117-2359 ns.🤖 Generated with Claude Code
https://claude.ai/code/session_01DDhgJqxUsdmWPKh6pKrV5A