Audio: revert clamps superseded by INT32 ramp overflow root-cause fix#153
Merged
Audio: revert clamps superseded by INT32 ramp overflow root-cause fix#153
Conversation
This reverts d9e6058. The 14.6% bilinear-warp overshoot it documented is real, but the only reason it produced audible square-wave clipping was that voice samples were being driven to int16 max — and that was the INT32 envelope-mixer ramp wrap, fixed at the root in e0160c0 (audio_envmix_ramp_overflow_2026-05-08). With the ramp clamp in place, voices no longer peak ±32767 (the e0160c0 verification shows L peak dropping 32767→27720 on the same scenario). The biquad's worst-case input is now well under the rails, the discrete step-response overshoot stays inside int16, and clampS16FromDouble never flat-tops. DK win-pose growl plays cleanly without the kHeadroom factor — verified manually in the worktree. Recovers ~1.4 dB of headroom on the output-filtered path that was previously thrown away to mask a different bug. The bug doc is kept and marked SUPERSEDED; the discrete-IIR step- response analysis remains a useful reference for future near-Nyquist filter work. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This reverts d30a967, plus the dependent code added by 0297c35 ("Fix reverb wet bus invisible to FX delay-line save/load") and 5432b06's scoped fixed-bus clear in aClearBufferImpl. d30a967 reintroduced int32 wet/dry bus accumulators (dry_left_acc[], wet_left_acc[], etc.) plus the fixed_bus_acc / sample_read / sample_write / sample_add infrastructure to dodge "loud SFX bus clipping." f0e2402 had previously removed the same int32-accumulator approach (originally introduced in 69fe2a7) on the grounds that it diverges from canonical N64 RSP behavior — both OOT (Ship of Harkinian) and SM64 reference ports use a saturating int16 DMEM bus with per-voice clamp16, not a wide accumulator clamped at the final stage. The "loud SFX bus clipping" symptom that motivated d30a967 was actually the same INT32 envelope-mixer ramp wrap fixed in e0160c0 manifesting through a different path: the wrapped ramp pushed individual voice contributions to ±32767, and stacking several of those on the saturating bus produced the cumulative-clipping artifact. With the ramp clamp in place, per-voice contributions stay in their intended range, and Sector Z multi-fighter brawls play cleanly against the canonical s16 saturating bus — verified manually in the worktree. 0297c35's aLoadBufferImpl/aSaveBufferImpl mirror logic existed only to keep the int32 bus state coherent with the s16 DMEM region across the FX delay-line DRAM round-trip. With the int32 accumulators gone, the s16 region is the only state and memcpy preserves it correctly; the mirror code is removed. Reverb is still audible because the bus path now matches what it was before the d30a967/0297c35 layering. 5432b06's scoped fixed-bus clear in aClearBufferImpl is removed for the same reason — there are no per-bus accumulators left to clear. The 14.5 kHz Butterworth output low-pass filter that 5432b06 also introduced is unrelated (anti-imaging concern, separate code path) and is kept. Net: 171 lines removed, audio path restored to the simpler RSP- faithful clamp16 saturating bus. The two superseded bug docs (audio_output_filter_overshoot_clip_2026-04-29 and audio_int32_acc_save_load_bridge_2026-04-29) are kept and marked SUPERSEDED so the discovery history remains traceable. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The submodule loop only iterated libultraship / torch / decomp; Battle- ShipYard (the modkit that builds BattleShip.fromsource.o2r) was registered in .gitmodules but never cloned into the worktree, which broke cmake configure on any branch where CMakeLists.txt's Battle-ShipYard/cmake/RelocData.cmake include is reachable. The existing skip-if-not-configured guard makes this safe on older base branches that don't yet track the submodule. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
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.
Summary
Two audio "fix" commits turned out to be band-aids over the int32 envelope-mixer ramp wrap that #108 / e0160c0 (
audio_envmix_ramp_overflow_2026-05-08) addressed at the root. With the int32→int64 ramp clamp in place, voice samples no longer peak ±32767, so the downstream symptoms those band-aids were chasing don't materialize. Both reverts have been verified manually against the original fail conditions.kHeadroom = 0.85.port/audio/mixer.c.The two superseded bug docs are kept and marked
SUPERSEDED 2026-05-08so the discovery history (and the discrete-IIR step-response analysis) remains traceable. The 14.5 kHz Butterworth output LPF that 5432b06 also introduced is unrelated (anti-imaging filter for content near Nyquist) and is kept.Bonus: the same branch fixes
scripts/new-worktree.shto clone theBattle-ShipYardsubmodule alongside the others, which the original loop missed.Test plan
🤖 Generated with Claude Code