Game: dark flashing textures, glitched Link, jumpy camera and persistent stutter while in the Kakariko Graveyard (RADV, RX 9070 XT) #1890

Open
opened 2026-11-20 17:41:46 +00:00 by nidrix_glandda · 9 comments

What happens

Standing in Kakariko Graveyard (entered from Kakariko Village, Graveyard_Adult), four things start together and do not stop until I leave the area:

  • frame rate collapses and stays low (persistent stutter, not a short load hitch),
  • Link's model glitches (parts of him deform/disappear),
  • textures on objects flash dark/black for short periods,
  • the camera is jumpy while turning.

The same cluster of symptoms happens on the path up to the graveyard in Kakariko Village (matches #1406). Leaving the area recovers it.

Builds it happens on

So it is base-main behaviour, not caused by any of those PRs.

Ruled out by switch A/Bs

HOS_COLLECT_VECTOR=0, RX_VK_DESC_SPLIT=0, HOS_TEXELS_KEEP_CORE=1, RX_VK_TEX_MOVE=1, RX_VK_WB_MAX_KB=0 + RX_VK_WB_DIFF=0, HOS_NATIVE_ZSTD=2, HOS_FPS_STEP=0, RX_VK_MESH_CONV=0. None removes the cluster.

Evidence from the logs

  • vk: texture #883 0x10002fe8e00 3440x1440 NVN 0x34: the GPU wrote its storage (submission 51276): uploaded again
  • vk: upload over a drawn image: #883 3440x1440 … at present 7325 (texture frame 7326), finalized, 0 KB dirty in 0 ranges, a GPU buffer wrote it
  • 26 upload over a drawn image events and 40 uploaded again events in one Graveyard run, with 194 textures retired.
  • During the episodes: main_cpu 240–280 ms per frame, game_ms up to 346 ms, main_blocked up to 345 ms, renderer submit/draw in single-digit ms, main_vcore ~0.03 ms — i.e. the frame is dominated by the game's own work/waits while the renderer is idle.
  • One run crashed at the transition: CRASH: exception 0xc0000005 (read 0x1100, guest 0xfffff00000001100) at /usr/lib/libgcc_s.so.1+0x26617 (old base build; not reproduced on pure main in 4 automated round trips).

Reproducer

A headless loop built on the repo's control bridge (scripts/longshot_ctl.py): launch with --profile test and a copied save (starts in Graveyard_Adult), press into play, then teleport Kakariko ↔ Graveyard with the debug teleport; sample HOS_SLOW_FRAMES, RX_VK_FRAME_TRACE, the log's texture lines, and screenshots after each arrival. No player input needed. Ask and I'll paste the script.

Separate regression found on this GPU (not the cause of the above)

#1846 (RX_VK_MESH_CONV, on by default) removes the vegetation entirely on this machine (RDNA4, no mesh shaders: the game's mesh draws run the compute fallback). RX_VK_MESH_CONV=0 brings it back. Reported here because it was found while testing; it is independent of the Graveyard cluster.

Environment

CachyOS (Arch), kernel 7.3.1, Mesa 26.3.1 RADV (GFX1201), Radeon RX 9070 XT 16 GB, Ryzen 5 5600X (12 threads), 32 GB RAM, 3440x1440, V-Sync Auto, frame generation off, Threaded rendering Auto (render + recorder threads).

Note on the automated captures

The teleport/dwell captures above come from the repo's control bridge with --profile test, which sets HOS_HEADLESS=1 (offscreen). docs/dev/reference/profiling.md says not to take frame-rate numbers from headless/profiled runs, so treat those fps figures as relative signals between builds/settings, not as the in-window fps. The loading re-entry (Sequence_IsLoadingBanc), the texture churn lines, the crash and the screenshots are unaffected by the window mode. A windowed HOS_PROFILE=1 run, as performance-report.md describes, is what a maintainer would want for the definitive numbers.

### What happens Standing in **Kakariko Graveyard** (entered from Kakariko Village, `Graveyard_Adult`), four things start **together** and do **not** stop until I leave the area: - frame rate collapses and stays low (persistent stutter, not a short load hitch), - Link's model glitches (parts of him deform/disappear), - textures on objects flash dark/black for short periods, - the camera is jumpy while turning. The same cluster of symptoms happens on the path up to the graveyard in Kakariko Village (matches #1406). Leaving the area recovers it. ### Builds it happens on - **current `main` `6d25da8c`** built from source, no local changes (pure main), - `89c3d21` + #1646, - that plus Tier 1 (#1827, #1870, #1776, #1851, #936), #1875, #1810, #941. So it is **base-main behaviour**, not caused by any of those PRs. ### Ruled out by switch A/Bs `HOS_COLLECT_VECTOR=0`, `RX_VK_DESC_SPLIT=0`, `HOS_TEXELS_KEEP_CORE=1`, `RX_VK_TEX_MOVE=1`, `RX_VK_WB_MAX_KB=0` + `RX_VK_WB_DIFF=0`, `HOS_NATIVE_ZSTD=2`, `HOS_FPS_STEP=0`, `RX_VK_MESH_CONV=0`. None removes the cluster. ### Evidence from the logs - `vk: texture #883 0x10002fe8e00 3440x1440 NVN 0x34: the GPU wrote its storage (submission 51276): uploaded again` - `vk: upload over a drawn image: #883 3440x1440 … at present 7325 (texture frame 7326), finalized, 0 KB dirty in 0 ranges, a GPU buffer wrote it` - 26 `upload over a drawn image` events and 40 `uploaded again` events in one Graveyard run, with 194 `textures retired`. - During the episodes: `main_cpu` 240–280 ms per frame, `game_ms` up to 346 ms, `main_blocked` up to 345 ms, renderer submit/draw in single-digit ms, `main_vcore` ~0.03 ms — i.e. the frame is dominated by the game's own work/waits while the renderer is idle. - One run crashed at the transition: `CRASH: exception 0xc0000005 (read 0x1100, guest 0xfffff00000001100) at /usr/lib/libgcc_s.so.1+0x26617` (old base build; not reproduced on pure main in 4 automated round trips). ### Reproducer A headless loop built on the repo's control bridge (`scripts/longshot_ctl.py`): launch with `--profile test` and a copied save (starts in `Graveyard_Adult`), press into play, then teleport Kakariko ↔ Graveyard with the debug teleport; sample `HOS_SLOW_FRAMES`, `RX_VK_FRAME_TRACE`, the log's texture lines, and screenshots after each arrival. No player input needed. Ask and I'll paste the script. ### Separate regression found on this GPU (not the cause of the above) #1846 (`RX_VK_MESH_CONV`, on by default) **removes the vegetation entirely** on this machine (RDNA4, no mesh shaders: the game's mesh draws run the compute fallback). `RX_VK_MESH_CONV=0` brings it back. Reported here because it was found while testing; it is independent of the Graveyard cluster. ### Environment CachyOS (Arch), kernel 7.3.1, Mesa 26.3.1 RADV (GFX1201), Radeon RX 9070 XT 16 GB, Ryzen 5 5600X (12 threads), 32 GB RAM, 3440x1440, V-Sync Auto, frame generation off, Threaded rendering Auto (render + recorder threads). ### Note on the automated captures The teleport/dwell captures above come from the repo's control bridge with `--profile test`, which sets `HOS_HEADLESS=1` (offscreen). `docs/dev/reference/profiling.md` says not to take frame-rate numbers from headless/profiled runs, so treat those fps figures as relative signals between builds/settings, not as the in-window fps. The loading re-entry (`Sequence_IsLoadingBanc`), the texture churn lines, the crash and the screenshots are unaffected by the window mode. A windowed `HOS_PROFILE=1` run, as [performance-report.md](https://git.magicbean.dev/MagicBeanDev/Longshot/src/branch/main/docs/user/performance-report.md) describes, is what a maintainer would want for the definitive numbers.
Member

Possibly related, for context:

  • issue #565 Game: Longshot , periodic stutter in kakariko village — Kakariko-area stutter, but another GPU and a less extensive symptom cluster
  • issue #1268 Game: 200-470 ms stutters when many new textures arrive: the main thread waits ~0.3 s on a GPU fence after a frame of uploads (RTX 3090, Linux) — Texture-upload-associated stutter investigation; NVIDIA and different measured mechanism
  • issue #1696 Game: Frame-time spikes while walking in circles near Kakariko bridge in Hyrule Field — Nearby Kakariko-region stutter on Linux/RADV, but a different location and trigger
  • PR #1846 perf(vk): the vegetation's indirect mesh draws keep their passes: their arguments converted at the game's barriers (RX_VK_MESH_CONV) — Merged mesh conversion change; report separately says its fallback removes vegetation
Possibly related, for context: - issue #565 Game: Longshot , periodic stutter in kakariko village — Kakariko-area stutter, but another GPU and a less extensive symptom cluster - issue #1268 Game: 200-470 ms stutters when many new textures arrive: the main thread waits ~0.3 s on a GPU fence after a frame of uploads (RTX 3090, Linux) — Texture-upload-associated stutter investigation; NVIDIA and different measured mechanism - issue #1696 Game: Frame-time spikes while walking in circles near Kakariko bridge in Hyrule Field — Nearby Kakariko-region stutter on Linux/RADV, but a different location and trigger - PR #1846 perf(vk): the vegetation's indirect mesh draws keep their passes: their arguments converted at the game's barriers (RX_VK_MESH_CONV) — Merged mesh conversion change; report separately says its fallback removes vegetation
Author

More evidence from an automated dwell in the Graveyard (no teleport, camera turned with the right stick, longshot-loop --dwell on current main + the PR stack with RX_VK_MESH_CONV=0):

  • sustained 10–16 fps while standing there (2880 presents, 10.0 fps; 1289 draws, 137 dispatches, 170 passes a frame), not a transient load tail;
  • the bridge log repeats control: state play -> loading and back to play within a frame or two: frames 2477, 2590, 2824, … — with 0.0 draws a frame at the transition and 1819.7 / 2139.0 / 1942.5 draws a frame right after. It looks like the game keeps re-entering its area/scene setup rather than doing one load;
  • slow-frame samples during it: ARes_Loading_0 17%, ARes_Loading_1 21%, ParallelDecompThread 17%, OptimizeChangeBanc 17%, while the main thread waits;
  • screenshots during the episode were checked; one showed Link with his arms out, but that may be an idle animation, so it is not counted as evidence. The dark/flashing textures, glitched Link model and jumpy camera are the player's reports from normal play.

So the four symptoms line up with the repeated scene re-entry: animation/texture state and camera on each re-entry, plus the frame rate (loaders + replay). No failed/missing asset line is logged (not found 0), and no map change is logged, so it is not a missing romfs file.

More evidence from an automated dwell in the Graveyard (no teleport, camera turned with the right stick, `longshot-loop --dwell` on current main + the PR stack with `RX_VK_MESH_CONV=0`): - sustained **10–16 fps** while standing there (2880 presents, 10.0 fps; 1289 draws, 137 dispatches, 170 passes a frame), not a transient load tail; - the bridge log repeats `control: state play -> loading` and back to play within a frame or two: frames 2477, 2590, 2824, … — with **0.0 draws a frame** at the transition and 1819.7 / 2139.0 / 1942.5 draws a frame right after. It looks like the game keeps re-entering its area/scene setup rather than doing one load; - slow-frame samples during it: `ARes_Loading_0` 17%, `ARes_Loading_1` 21%, `ParallelDecompThread` 17%, `OptimizeChangeBanc` 17%, while the main thread waits; - screenshots during the episode were checked; one showed Link with his arms out, but that may be an idle animation, so it is not counted as evidence. The dark/flashing textures, glitched Link model and jumpy camera are the player's reports from normal play. So the four symptoms line up with the repeated scene re-entry: animation/texture state and camera on each re-entry, plus the frame rate (loaders + replay). No failed/missing asset line is logged (`not found` 0), and no map change is logged, so it is not a missing romfs file.
Author

Area control, same build/run method (longshot-loop --dwell, RX_VK_MESH_CONV=0): Hyrule Field (adult) 58.2-60.2 fps over four samples with 0 state play -> loading transitions; Kakariko Graveyard 10-16 fps with loader threads busy and Sequence_IsLoadingBanc re-entering. HOS_OCB_HOLD_MS=10000 removed the extra re-entries in one dwell (only the two teleports remained) but fps stayed 11.8, so the preload hold alone is not the fix. Next: HOS_CHAIN across a Graveyard dwell to name each re-entry wait.

Area control, same build/run method (`longshot-loop --dwell`, `RX_VK_MESH_CONV=0`): Hyrule Field (adult) 58.2-60.2 fps over four samples with 0 `state play -> loading` transitions; Kakariko Graveyard 10-16 fps with loader threads busy and `Sequence_IsLoadingBanc` re-entering. `HOS_OCB_HOLD_MS=10000` removed the extra re-entries in one dwell (only the two teleports remained) but fps stayed 11.8, so the preload hold alone is not the fix. Next: `HOS_CHAIN` across a Graveyard dwell to name each re-entry wait.
Author

Attached the HOS_PROFILE=1 profile of a Graveyard run that ends in the crash (1.4 MB zip): longshot-profile-20261120-194606.zip

Build: current main 6d25da8c + Tier 1 + #1875/#1810/#941, RX_VK_MESH_CONV=0 (vegetation regression of #1846 avoided). The run crashes with 0xc0000005 during the Kakariko -> Graveyard teleport, as before. Before the crash, report.md says:

  • 1867 frames in 24 s, 76.6 fps average, but the 10 s windows go 133.3 -> 32.7 -> 20.7 fps; 47 frames of 50 ms or more, p99 84 ms, worst 727 ms;
  • 43 of the 47 stutters are dominated by main_blocked_ms (94.8 ms average of a 106 ms stutter); draw translation 3, shader translation 1;
  • no game thread running in 59 % of the stutters' instants (48 % otherwise): the frame waits outside the game's CPU work; 7 of the run's 33 long virtual-core waits overlap a stutter;
  • steady frames: game 2.5 ms, renderer 0.1 ms, GPU wait 0.0 ms.

trace.json.gz (Perfetto) and speedscope.json (flame graphs) are in the zip. RX_VK_VRAM_MB=12000 was tested too: no difference in fps or retires, so it is not the texture budget.

Attached the `HOS_PROFILE=1` profile of a Graveyard run that ends in the crash (1.4 MB zip): [longshot-profile-20261120-194606.zip](https://git.magicbean.dev/attachments/5fe08741-c854-46a1-bef2-102bf777e8bf) Build: current main `6d25da8c` + Tier 1 + #1875/#1810/#941, `RX_VK_MESH_CONV=0` (vegetation regression of #1846 avoided). The run crashes with `0xc0000005` during the Kakariko -> Graveyard teleport, as before. Before the crash, `report.md` says: - 1867 frames in 24 s, 76.6 fps average, but the 10 s windows go 133.3 -> 32.7 -> 20.7 fps; 47 frames of 50 ms or more, p99 84 ms, worst 727 ms; - 43 of the 47 stutters are dominated by `main_blocked_ms` (94.8 ms average of a 106 ms stutter); draw translation 3, shader translation 1; - **no game thread running in 59 % of the stutters' instants** (48 % otherwise): the frame waits outside the game's CPU work; 7 of the run's 33 long virtual-core waits overlap a stutter; - steady frames: game 2.5 ms, renderer 0.1 ms, GPU wait 0.0 ms. `trace.json.gz` (Perfetto) and `speedscope.json` (flame graphs) are in the zip. `RX_VK_VRAM_MB=12000` was tested too: no difference in fps or retires, so it is not the texture budget.
Author

Confirmed fixed by the new build: v3.0.0-n627-g93e8237d = nightly-627 (afc4d1ad, which already has #792/#936/#1866/#1781/#1809/#1873/#1727/#327/#1646) + #1642 (HOS_BG_THROTTLE, on by default), #1870, #1893, #1914, #1909/#1911 (off), with RX_VK_MESH_CONV=0 for the #1846 vegetation regression.

In the Graveyard now: steady 60 fps, no loading re-entry, no dark flashes, no Link/camera glitch. The likely contributors are #1642 (the OptimizeChange throttle, on by default) and the now-merged renderer/thread work (#792 image reaper, #1866 recorder parts, #1781/#1809). Closing; reopen if it comes back on a newer base.

Confirmed fixed by the new build: `v3.0.0-n627-g93e8237d` = nightly-627 (`afc4d1ad`, which already has #792/#936/#1866/#1781/#1809/#1873/#1727/#327/#1646) + #1642 (`HOS_BG_THROTTLE`, on by default), #1870, #1893, #1914, #1909/#1911 (off), with `RX_VK_MESH_CONV=0` for the #1846 vegetation regression. In the Graveyard now: steady 60 fps, no loading re-entry, no dark flashes, no Link/camera glitch. The likely contributors are #1642 (the OptimizeChange throttle, on by default) and the now-merged renderer/thread work (#792 image reaper, #1866 recorder parts, #1781/#1809). Closing; reopen if it comes back on a newer base.
Author

Reopening: on v3.0.0-n631-g67220bf2 (nightly-631 e0ef1b57 + #1642/#1870/#1893/#1914/#1909/#1911, RX_VK_MESH_CONV=0) the persistent Graveyard/Kakariko state is gone (60 fps, no loading re-entry), but the problem is not fully fixed:

  • a visible 1–3 s hiccup still happens in some areas, together with dark/flashing textures and the jumpy camera, as before;
  • unlike before it now recovers by itself, so the self-feeding stop has been removed;
  • the rumble/vibration problem is gone on this build (but Vibration was also Off in the session, so not proof).

So this looks like the residual neighbour-Banc/loading burst rather than the stuck state: #1642 (HOS_BG_THROTTLE, on) and the merged #792/#1866/#1881 work removed the part that stayed broken. Leaving this open for the transient; next thing to try is HOS_BG_THROTTLE variants and a windowed HOS_PROFILE=1 capture at the moment of the 1–3 s hiccup, now that the PC is available.

Reopening: on `v3.0.0-n631-g67220bf2` (nightly-631 `e0ef1b57` + #1642/#1870/#1893/#1914/#1909/#1911, `RX_VK_MESH_CONV=0`) the **persistent** Graveyard/Kakariko state is gone (60 fps, no loading re-entry), but the problem is not fully fixed: - a visible **1–3 s hiccup** still happens in some areas, together with **dark/flashing textures** and the jumpy camera, as before; - unlike before it now **recovers by itself**, so the self-feeding stop has been removed; - the rumble/vibration problem is gone on this build (but Vibration was also Off in the session, so not proof). So this looks like the residual neighbour-Banc/loading burst rather than the stuck state: #1642 (`HOS_BG_THROTTLE`, on) and the merged #792/#1866/#1881 work removed the part that stayed broken. Leaving this open for the transient; next thing to try is `HOS_BG_THROTTLE` variants and a windowed `HOS_PROFILE=1` capture at the moment of the 1–3 s hiccup, now that the PC is available.
Author

Deterministic reproducer added: a script that teleports to a map, holds the left stick forward for N seconds and samples the bridge fps between chunks (headless, --profile test, own save copy). Results on the current build (v3.0.0-n631-g67220bf2 = nightly-631 + #1642/#1870/#1893/#1914/#1909/#1911, RX_VK_MESH_CONV=0):

  • Kokiri Forest, running straight 8 s: steady 59.1-60.2 fps.
  • Kakariko Village, same: one dip to 38.4 fps near the start, then 60.
  • Graveyard, same: 60 -> 35.5 -> 19.7 -> 60 -> 11.8 -> 51 (80 % drop max-to-min).

The frame trace gives hiccup clusters of 9-12 frames over 0.8-0.9 s, peak frames 240-288 ms in the Graveyard (447 ms in Kakariko) - matching the seen 1-3 s hiccup with the dark flashing textures, which now recovers. Next: A/B HOS_BG_THROTTLE (default 300:3000 vs 0 vs 100:1000) with this same straight-line run, and a windowed HOS_PROFILE=1 at the hiccup.

Deterministic reproducer added: a script that teleports to a map, holds the left stick forward for N seconds and samples the bridge fps between chunks (headless, `--profile test`, own save copy). Results on the current build (`v3.0.0-n631-g67220bf2` = nightly-631 + #1642/#1870/#1893/#1914/#1909/#1911, `RX_VK_MESH_CONV=0`): - Kokiri Forest, running straight 8 s: steady 59.1-60.2 fps. - Kakariko Village, same: one dip to 38.4 fps near the start, then 60. - Graveyard, same: 60 -> 35.5 -> **19.7** -> 60 -> **11.8** -> 51 (80 % drop max-to-min). The frame trace gives hiccup clusters of 9-12 frames over 0.8-0.9 s, peak frames 240-288 ms in the Graveyard (447 ms in Kakariko) - matching the seen 1-3 s hiccup with the dark flashing textures, which now recovers. Next: A/B `HOS_BG_THROTTLE` (default 300:3000 vs 0 vs 100:1000) with this same straight-line run, and a windowed `HOS_PROFILE=1` at the hiccup.
Author

Windowed measurement on v3.0.0-n679-g0f554e33 = nightly-679 (b62932a9, includes #1946 which restores the compute fallback vegetation, #2008, #1933, #2007, #2012, #1574, #1914, #1870) + #1642, #1893, #1909, #2036, #2047. No RX_VK_MESH_CONV override; vegetation is back on default.

8 s straight runs, real window (own user dir + save copy):

  • Graveyard: 57.9-61.7 fps
  • Kakariko: 56.7-60.9 fps

No sampled fps dip, but the frame trace still shows short clusters of >33 ms frames: 9 frames / 0.83 s (peak 284 ms) and 3 frames / 0.58 s (peak 413 ms) in the Graveyard; 17 frames / 1.13 s (peak 260 ms) in Kakariko. So the remaining 1-3 s hiccup with dark flashes is still there but rarer/shorter. Earlier headless figures were misleading (profiling.md says not to use them for frame rate).

Windowed measurement on `v3.0.0-n679-g0f554e33` = nightly-679 (`b62932a9`, includes #1946 which restores the compute fallback vegetation, #2008, #1933, #2007, #2012, #1574, #1914, #1870) + #1642, #1893, #1909, #2036, #2047. No `RX_VK_MESH_CONV` override; vegetation is back on default. 8 s straight runs, real window (own user dir + save copy): - Graveyard: 57.9-61.7 fps - Kakariko: 56.7-60.9 fps No sampled fps dip, but the frame trace still shows short clusters of >33 ms frames: 9 frames / 0.83 s (peak 284 ms) and 3 frames / 0.58 s (peak 413 ms) in the Graveyard; 17 frames / 1.13 s (peak 260 ms) in Kakariko. So the remaining 1-3 s hiccup with dark flashes is still there but rarer/shorter. Earlier headless figures were misleading (profiling.md says not to use them for frame rate).
Author

Windowed area sweep on v3.0.0-n679-0f554e33 (nightly-679 + #1642/#1893/#1909/#2036/#2047), own user dir + save copy, 6 s straight run per area, both ages, 18 areas:

  • every adult area sits at 60 fps (worst sample 56.1 in Lost Woods);
  • child: also ~60 except Castle Town Market, child: median 51.0 fps (48.2-57.0) and Kakariko child dipped to 52.1 once.

The run's frame trace (15,299 frames / 266 s, teleport loads included) has 486 frames over 33 ms (3.2 %), worst 1,144 ms, with nine 0.5-1.3 s clusters. Many coincide with the arrival load after a teleport, but the Market-child stretch holds ~51 fps for its whole 6 s instead of 60. Screenshots and the trace are local; happy to attach the trace if wanted.

Windowed area sweep on `v3.0.0-n679-0f554e33` (nightly-679 + #1642/#1893/#1909/#2036/#2047), own user dir + save copy, 6 s straight run per area, both ages, 18 areas: - every adult area sits at 60 fps (worst sample 56.1 in Lost Woods); - child: also ~60 except **Castle Town Market, child: median 51.0 fps (48.2-57.0)** and Kakariko child dipped to 52.1 once. The run's frame trace (15,299 frames / 266 s, teleport loads included) has 486 frames over 33 ms (3.2 %), worst 1,144 ms, with nine 0.5-1.3 s clusters. Many coincide with the arrival load after a teleport, but the Market-child stretch holds ~51 fps for its whole 6 s instead of 60. Screenshots and the trace are local; happy to attach the trace if wanted.
Sign in to join this conversation.
No milestone
No project
No assignees
2 participants
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
MagicBeanDev/Longshot#1890
No description provided.