fix(vk): texture noise when system RAM runs low: a staging arena of the size needed when 64 MB can't be pinned #1514

Merged

rakefire-impala merged 1 commit from fix/texture-noise-low-ram-staging-fallback into main

2026-11-10 14:47:57 +00:00

rakefire-impala commented

2026-11-10 05:44:06 +00:00

(Migrated from github.com)

Related: #993. Part of splitting up #1023.

Problem

On an RX 5500M laptop (4 GB VRAM, 16 GB RAM), the game's pools were imported (about 2 GB pinned), only 2.9 GB of RAM was available, and Windows was paging. texture_frame had given the idle arenas back, as it does after 30 s on a card of 6 GB or less, and making a new 64 MB one failed:

vk: staging: an arena of 64 MB for 36 KB could not be made (1 arenas)
vk: memory heaps: heap 0 (device): ... used 2972 of 2972 MB; heap 1 (system): ... used 3661 of 7206 MB; heap 2 (device): ... used 198 of 198 MB;
vk: BC3 texture #... its upload failed (no staging memory for its ASTC blocks): tried again in 30 frames

The system heap's budget had room, but pinning a new 64 MB block failed under RAM pressure. Every upload failed until an arena came free.

Change

When a 64 MB arena can't be made, staging() tries one of the size the upload needs, rounded up to 1 MB. Both outcomes are logged once in a while: vk: staging: no 64 MB arena could be made (N kept); one of the size asked for instead, or ...; the upload waits.
A buffer that is made but not mapped is now destroyed instead of leaked.
texture_frame gives the small arenas back as it does the others. Upstream's arena timing log is kept.

Who it affects

Only a machine where a 64 MB staging arena can't be pinned: low free RAM, no resizable BAR. Everywhere else staging() makes the same 64 MB arenas as before, and the new path never runs.

Tests

tests/test_staging_arena_fallback.py (new) checks the order of the attempts, the cleanup and the log line. It fails on main without this change.

Tested on


Tested on	RX 5500M (RDNA1, 4 GB VRAM, 16 GB RAM), Windows 11, AMD driver; Hookshot main 0c412ad (v2.0.11-383) with these PRs, 2026-11-10
Result	In one session, `no 64 MB arena could be made ...; one of the size asked for instead` 8 times, 0 failed texture uploads, no noise

Without resizable BAR, a texture staging arena is 64 MB of system memory the driver must pin. On a 16 GB laptop with RAM nearly gone, no 64 MB arena could be made, even for a 36 KB upload, so texture uploads failed and drew as noise. This falls back to an arena of the size the upload needs. Related: #993. Part of splitting up #1023. ## Problem On an RX 5500M laptop (4 GB VRAM, 16 GB RAM), the game's pools were imported (about 2 GB pinned), only 2.9 GB of RAM was available, and Windows was paging. `texture_frame` had given the idle arenas back, as it does after 30 s on a card of 6 GB or less, and making a new 64 MB one failed: ``` vk: staging: an arena of 64 MB for 36 KB could not be made (1 arenas) vk: memory heaps: heap 0 (device): ... used 2972 of 2972 MB; heap 1 (system): ... used 3661 of 7206 MB; heap 2 (device): ... used 198 of 198 MB; vk: BC3 texture #... its upload failed (no staging memory for its ASTC blocks): tried again in 30 frames ``` The system heap's budget had room, but pinning a new 64 MB block failed under RAM pressure. Every upload failed until an arena came free. ## Change - When a 64 MB arena can't be made, `staging()` tries one of the size the upload needs, rounded up to 1 MB. Both outcomes are logged once in a while: `vk: staging: no 64 MB arena could be made (N kept); one of the size asked for instead`, or `...; the upload waits`. - A buffer that is made but not mapped is now destroyed instead of leaked. - `texture_frame` gives the small arenas back as it does the others. Upstream's arena timing log is kept. ## Who it affects Only a machine where a 64 MB staging arena can't be pinned: low free RAM, no resizable BAR. Everywhere else `staging()` makes the same 64 MB arenas as before, and the new path never runs. ## Tests `tests/test_staging_arena_fallback.py` (new) checks the order of the attempts, the cleanup and the log line. It fails on main without this change. ## Tested on | | | |---|---| | Tested on | RX 5500M (RDNA1, 4 GB VRAM, 16 GB RAM), Windows 11, AMD driver; Hookshot main 0c412ad (v2.0.11-383) with these PRs, 2026-11-10 | | Result | In one session, `no 64 MB arena could be made ...; one of the size asked for instead` 8 times, 0 failed texture uploads, no noise |