Skip to content

vkQuake 1.37.0 demo1 1080p hangs the GPU on Apple M1 with r_oit 0 (Unmapped read faults); r_oit 1 is clean #16

Open

Description

@joshuaswarren

Summary

vkQuake 1.37.0 demo1 at 1920x1080 (windowed) hangs the GPU on an Apple M1 (G13G) with r_oit 0, the vkQuake default. Each hang is a GPU timeout with an Unmapped read fault. The same run with r_oit 1 completes without any timeout.

Setup

  • Apple M1 (13-inch), kernel 7.1.12-2-20261013.18-sep-ARCH.
  • System Mesa 1:26.2.4-1, Vulkan driver vulkan-asahi-omarchy 1:26.2.4.git69bde15340-2. No agx: exact division, log and sin/cos for compute; graphics keeps the base math #14 code is involved.
  • vkquake 1.37.0-2, sdl3 3.4.16-1, Wayland session, shareware pak0.pak (sha256 35a9c55e5e5a284a159ad2a62e0e8def23d829561fe2f54eb402dbc0a9a946af), demo1.
  • Command: vkquake -basedir <dir> -width 1920 -height 1080 -window -nosound -nostartdemos +vid_vsync 0 +host_maxfps 200 <extra cvars> +timedemo demo1. Each run was killed at the first GPU timeout or after 120 s.

Results (GPU timeout lines added in dmesg per run)

Extra cvars New timeouts
none (defaults) 4
+r_rtshadows 0 2
+r_waterwarpcompute 1 23
+r_oit 1 0
+r_oit 1 +r_waterwarpcompute 1 +r_rtshadows 0 +vid_fsaa 0 +r_scale 1 +gamma 0.9 +contrast 1.4 0

So r_oit 0 triggers it. r_waterwarpcompute 1 with r_oit 0 is the worst case.

Fault samples

Defaults, timeout 47: unit: UL1C(0x3), address: 0x11fffbd400, read: true, reason: Unmapped, vm_slot: 0xf.

r_waterwarpcompute 1, timeout 77: unit: dVDM_CSD, address: 0x37df000000, read: true, reason: Unmapped, vm_slot: 0x22, one pending event [0:16] flags=15.

All 78 faults in the log read reason: Unmapped.

Not measured

Activity

  1. joshuaswarren commented on Nov 5, 2026

    @joshuaswarren
    Author

    Status 2026-11-05 (w7K): no retest yet. A retest needs an M1 run that goes through the GPU ticket queue, and it has not started. Staying open until there is data: either a retest on the current head, or a close with the result.

  2. joshuaswarren commented on Nov 5, 2026

    @joshuaswarren
    Author

    Status, 2026-11-05 (single run, 04:05 CST, 13-inch M1 host).

    Result: demo1 at 1920x1080 with r_oit 0 hung the GPU again. The demo played into the first pickups, then the log stops at "Reallocating dynamic vertex buffers (512 KB)". No result line printed, and the r_oit readback did not print before the hang.

    Versions: kernel 7.1.12-2-20261013.18-sep-ARCH; mesa 1:26.2.4-1; vulkan-asahi-omarchy 1:26.2.4.git69bde15340-2; vkquake 1.37.0-2 (binary SHA256 a2924ed8...); driver library SHA256 40282811....

    Kernel journal for the same window: 8 GPU timeout events (Timeout count 0 to 7), each with reason: Unmapped, vm_slot: 0x37, and an Attempting recovery line. Event slots 77 and 79 are shown; a later event used slot 78. Excerpt, hostname removed:

    Nov 05 04:05:12.095739 <host> kernel: asahi 206400000.gpu: ** GPU timeout nya~!!!!! **
    Nov 05 04:05:12.095920 <host> kernel: asahi 206400000.gpu:   Event slot: 77
    Nov 05 04:05:12.096112 <host> kernel: asahi 206400000.gpu:   Timeout count: 0
                                                  vm_slot: 0x37,
                                                  reason: Unmapped,
    Nov 05 04:05:12.097731 <host> kernel: asahi 206400000.gpu:   Attempting recovery...
    Nov 05 04:05:12.098915 <host> kernel: asahi 206400000.gpu: ** GPU timeout nya~!!!!! **
    Nov 05 04:05:12.099074 <host> kernel: asahi 206400000.gpu:   Event slot: 79
    Nov 05 04:05:12.099253 <host> kernel: asahi 206400000.gpu:   Timeout count: 1
    Nov 05 04:05:12.115094 <host> kernel: asahi 206400000.gpu:   Timeout count: 7
    

    Limits:

    • One run. Not repeated.
    • The readback was queued after the timedemo command in that run, so its absence does not show the setting was wrong. The next revision queues it first.
    • The runner stopped before writing its post-run receipts, so the fault count comes from a separate journal read, not from the runner's output.
    • No r_oit 1 comparison was run in this session.
  3. joshuaswarren commented on Nov 5, 2026

    @joshuaswarren
    Author

    Follow-up to my comment linked above with a second run on the same 13-inch M1.

    Second run (rev 8 of the test packet, 2026-11-05). Same demo1 timedemo, same vkQuake 1.37.0 binary and pak0 hashes as the packet, mesa 1:26.2.4-1 and vulkan-asahi-omarchy 1:26.2.4.git69bde15340-2, after a clean reboot (new boot id). The runner started the game at 04:40:52 CST with timeout --kill-after=4 15, which sends SIGTERM at 15 seconds. The kernel logged eight GPU timeouts, all stamped 04:41:07 CST, which is 15 seconds after the start.

    The run log shows that the timedemo started ("Playing demo from demo1.dem"). The last game line is "Reallocating dynamic vertex buffers (512 KB)", the same line where the first run stopped. No result line was printed. The runner printed no summary: its summary file holds only the start line, and the watch ended with status unknown and no return code.

    The kernel log cannot show whether the GPU stopped during the demo or during the runner's termination at 15 seconds. The timeouts are at the bound, so I am not calling this a hang. I am reporting the fault sequence.

    Kernel log, abridged. Timestamps, host name, the ASCII art, the Unk lines, and most FaultInfo fields are removed. The lines below are from the first and last events:

    asahi 206400000.gpu: ** GPU timeout nya~!!!!! **
    asahi 206400000.gpu:   Event slot: 41
    asahi 206400000.gpu:   Timeout count: 0
    asahi 206400000.gpu:   Fault info: FaultInfo { ... address: 0x0, unit: DCMP, vm_slot: 0x17, read: true, reason: Unmapped, ... }
    asahi 206400000.gpu:   Pending events: [0:40] ... [1:9] ... [2:41] ...
    asahi 206400000.gpu:   Halt count: 1
    asahi 206400000.gpu:   Halted: 1
    asahi 206400000.gpu:   Attempting recovery...
      (the same block repeats for each event)
    asahi 206400000.gpu:   Event slot: 43
    asahi 206400000.gpu:   Timeout count: 7
    asahi 206400000.gpu:   Fault info: FaultInfo { ... address: 0x11fffbd400, unit: UL1C(0x3), vm_slot: 0x17, read: true, reason: Unmapped, ... }
    asahi 206400000.gpu:   Halt count: 8
    asahi 206400000.gpu:   Halted: 1
    asahi 206400000.gpu:   Attempting recovery...
    

    The eight events: event 1 is slot 41 at address 0x0 on unit DCMP with three pending events. Events 2 to 8 are slot 43 on unit UL1C, at addresses 0x11fffbc180, 0x11fffbc700, 0x11fffbc200, 0x11fffbd280 (twice), 0x11fffbd200, and 0x11fffbd400. Every event is reason: Unmapped, read: true, vm_slot: 0x17. Recovery was attempted after each. The issue body's "Defaults, timeout 47" sample has the same address and unit as event 8, with vm_slot: 0xf.

    First run (rev 5, 2026-11-05, 04:05 CST, earlier boot). Eight timeout events, reason Unmapped, on slots 77, 79 and 78, with vm_slot 0x37, and the same recovery pattern. The journal excerpt is in my comment linked above. The first run queued its readback after the timedemo, so the absence of its readback proves nothing. The first run also stopped at "Reallocating dynamic vertex buffers (512 KB)".

    Runner limits

    • Readback: the second run queued +r_oit 0 +r_oit before the timedemo. The readback did not print to the stdout log. The console file was not copied because the run stopped first, so I cannot say whether it printed to the console.
    • Summary: neither run produced a result or summary line.
    • No FPS or pass for either run.

    Not tested in this session

    • No r_oit 1 run in this session. The issue body reports that r_oit 1 completed with no timeouts on this hardware in an earlier session. These two runs do not repeat that comparison.
    • No other demo, no pak1, and no other Vulkan application on this driver.
    • No Mesa or kernel bisect. The executable in Chris Kearney's receipt (agx: exact division, log and sin/cos for compute; graphics keeps the base math #14 comment 6401786759, SHA256 3661f258...) does not match the pinned binary, and I have not tested it.
    • No measured FPS or frame times at any setting on this hardware.

    The vkQuake retest is stopped on this path, and we are not running it again.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions