Quake — PPSA01880: Linux bring-up and PCM audio investigation

[JJenkx]

Quake, PPSA01880 v01.009.000, Linux bring-up and audio validation, 2026-10-15.

Observed behavior

A native run on main ec1193dc3 plus #6029, #6476 and the PCM implementation in draft #5548 reaches the title/menu, then the rendered start map with HUD. It closes normally after an ordinary window-close request at about 191 seconds. A control with the same main, #6029 and #6476, but without PCM retention, also renders and closes normally. No complete playthrough, save/load round trip, or current listening verdict is established.

The same six scheduled button presses led to different scenes in the two runs. These are diagnostic startup/audio captures, not a matched-scene benchmark. No FPS improvement is claimed.

Open work

• fix(ngs2): retain PCM input when reporting decoded bytes #5548: experimental PCM ownership and decoded-byte accounting. Future input remains guest-owned until first decoding; retained bytes are counted before the guest may reuse them. Console decoder granularity, repeats and waveform-pointer timing remain unmeasured, so the PR stays a draft. This does not supersede earlier reports of static audio.
• fix(ngs2): stop stale rendering after waveform callbacks #6029: callback waveform replacement stops the stale rendering path; included in both native integrations above, with the original author's commit preserved in fix(ngs2): retain PCM input when reporting decoded bytes #5548.
• fix(build): preserve bundled validator type tables through exit #6476: build-applied SPIRV-Tools validator table-lifetime repair. Both runs use the official dependency gitlink, with this source patch, rather than a private dependency checkout. This is not a general title-shutdown fix.

The PCM draft is now rebased onto main fa9365ca7 (public head b21f78b004f82469b8bad0dca31e26e8bb57eb5b). Its full runtime build and 14 NGS2 tests pass on the refreshed base, including new stereo released-ring tests at 48 kHz and 44.1 kHz in both NGS2 library variants. A stereo byte-undercount fault fails both ownership tests; restoring the code passes with unchanged test executables. The native runs predate this last rebase; no native verdict is claimed for its newer graphics changes.

Audio evidence and limits

The control produced 29,698 stereo-float grains and the candidate 29,621 (256 frames per grain). Each accepted SDL float buffer matched the corresponding producer buffer byte-for-byte: 29,697 control buffers and 29,620 candidate buffers, with one unmatched final producer grain in each capture. Producer samples were finite in both.

For audio seconds 60–100, measured from the first producer record, the ratio of grain-boundary jump RMS to within-grain jump RMS was 10.0122 for the control and 1.0091 for the candidate. That is useful evidence of a boundary-concentration pattern, but the different rendered scenes prevent controlled quantitative attribution. Peaks exceeded 1.0 in both runs. Finite samples and producer/SDL matching do not establish speaker output, clipping elimination or PS5 equivalence. Capture interposition adds overhead.

Native source heads: control 4cd78bbbeecf646dbcc5f11c701deb8bbbb3c2cf, candidate 0cbfb31edf7c39b9ecfa0984afb15f13045e6202. Current-build NGS2 mappings and clean source were verified. Both runs restored the prior 64 MiB udmabuf limit and left no native game process running.

Machine: Ryzen 9 7950X, AMD Radeon RX 6650 XT, RADV Mesa 26.2.4, Arch Linux / kernel 7.2.9-arch1-1. NGS2 regressions are CPU tests. No proprietary assets, captured audio, screenshots, binaries or raw title logs are attached.

AI-assisted: yes.