Street Fighter 6 (PPSA02632): offline menus reached on main, severe slowdown

[embersas]

Status: Menus, offline

Street Fighter 6, PPSA02632. Tested on 20 October 2026 with unmodified main 525e299fc085c400de4a703c6b1ab105eefb50c7.

The game reaches the offline World Tour / Battle Hub / Fighting Ground mode-selection menu. Interactive fighting gameplay has not been tested, so this is a Menus result.

Environment

Item	Tested configuration
Host	Ryzen 9 3900X, 64 GB RAM
GPU	NVIDIA RTX 3080 Ti 12 GB, driver 610.88
OS	Windows 11 Pro 10.0.26300
Build	Release, supported WinLibs GCC 15.2.0 UCRT POSIX SEH r7
Title metadata	contentVersion 01.000.000, masterVersion 01.01

The relinker, all system libraries and both bundled title modules were freshly built/converted for this test. Deployed system libraries were checked against the build. No local source changes or additional PRs were needed to reach this menu. This does not establish what later gameplay requires.

Route to the offline menu

1. Launch the converted title with the matching runtime libraries. Let it populate the new compiler's cache, then restart with that cache.
2. Wait for Press any button and confirm. Upstream's tools/pad_input.py was also used for controller pulses during testing.
3. Online login displayed A communication error has occurred, code 20607-10005 R1183-0-0.
4. Dismiss the error. When offered Start the game in offline mode?, choose Yes.
5. The introductory sequence runs. Options reveals Cross: Skip; Cross skips it.
6. The three-panel mode-selection menu appears.

Current limitations

• Severe slowdown: a warmed restart recorded 13,033 shader-cache hits and zero misses at the title screen, while observed title rates were around 2-3 FPS. The later menu transition had long pauses and temporarily fell below 1 FPS, but it also compiled additional shaders. These observations are not a performance benchmark or a speedup claim. A clean performance trace is being collected separately.
• Online login failed: the error above appeared during this test route; its cause has not been diagnosed. The offered offline route worked.
• Normal close did not terminate one run within at least ten seconds: that process was stopped explicitly. A separate shutdown diagnosis has not been performed.
• Gameplay and rendering correctness remain unverified: introductory/attract sequences rendered, but no interactive fight was tested. Transient pauses resumed, so no permanent deadlock is claimed.

The original descriptor-Phi abort in #7459 did not recur in normal startup. Neither did the older storage-image write-watch abort during these bounded runs. This does not prove either failure family fixed, and it does not validate a particular PR.

Compatibility update: #7459 (comment)

This report contains sanitised observations only. Raw title logs, game data and shader captures remain private.

AI-assisted report; local build provenance, runtime-library hashes, cache counters and on-screen progress were checked.

[embersas]

Performance follow-up: warmed title-screen slowdown

Collected on 20 October 2026 using the same clean Release build, main 525e299fc085c400de4a703c6b1ab105eefb50c7. Local Ryzen 9 3900X, 64 GB RAM, RTX 3080 Ti 12 GB, NVIDIA 610.88, Windows 11 Pro 10.0.26300. Video output was 3840x2160.

The settled Press any button scene remains slow after a cache-warming run and restart. This is a short title-screen measurement; interactive fighting performance remains untested.

Capture and results

• Restarted with the previous run's shader and pipeline caches. No game input during capture. The title scene was visually confirmed before and after the window, at presents 825 and 943.
• APS5_TRACE_FRAMES=900:20, with APS5_TRACE_DIR as the destination. No other APS5_TRACE_* or APS5_PROFILE_* switches. Overlay and input helper disabled; application logs were not redirected.
• The SF6 process inherited the process-local SEM_FAILCRITICALERRORS setting so failed Windows module probes returned errors without modal Bad Image dialogs. No runtime files or system security settings were changed.
• The trace and its summary identify the expected clean commit (dirty=false), hardware and settings. Upstream tools/trace_digest.py parsed it successfully. No dropped events.

Measurement	Settled title, frames 900:20
Presents / window duration	20 / 2,493.214 ms
Rate over this window	approximately 8.02 presents/second
Frame time p50 / p90	124.581 / 139.435 ms
Frame time p99 / maximum	152.234 / 152.234 ms
Recorded GPU busy per present	21.454 ms

Selected summary/digest findings:

Scope	Inclusive ms per present
Driver.Packet.memory_wait	116.068
Vulkan.Dispatch.shader_resources	69.111
Resources.Complete	53.443

Queue worker 0x0 spent 83.879% of the window in the summary's packet/submission scopes. That includes waits and is not a CPU utilisation percentage. The scopes above overlap, can span multiple threads, and must not be added as independent frame costs. Upstream also documents that GPU timing can omit the final in-flight batches.

These measurements point towards CPU-side resource preparation and synchronisation waits as the next investigation area. They do not establish the older address-based host-import mechanism as the current cause or validate a particular PR.

Separate stall observation

An earlier capture at 1000:20 recorded 20 presents over 13,495.279 ms, with a 453.823 ms median and a 4,997.043 ms maximum frame. It also recorded substantial graphics-pipeline preparation (Vulkan.GraphicsPipelinePrebuild, 275.365 inclusive ms per present), and the title background visibly changed to a large black region while Press any button remained visible. No events were dropped.

That is a supporting diagnostic observation. The trace does not establish what caused the stall or black region. It needs repetition before treating either as a separate bug.

The two windows cover different moments, and the later run had additional cache warm-up. Neither is a controlled speed comparison, including against the earlier overlay/logged 2-3 FPS observations. The settled capture is only 20 presents over 2.49 seconds; its percentile values describe that small sample, not a long session.

This post contains sanitised measurements and digest findings only. Raw traces, title logs and captures remain private. No gameplay, rendering-correctness or speedup claim is made.

AI-assisted report; local trace summaries, upstream digests, capture settings and deployed runtime hashes were checked.