[wolfchimneyrock]

On Arm's own Vulkan driver (libmali) the game's memory grows by about 490 MB an hour, until it runs out of memory after a few hours (sooner on 4 GB boards). PanVK and Turnip don't grow.

Cause. A malloc tracker on the running game put every growing allocation inside libmali: 64 KB host-memory blocks allocated by vkCmdBeginRendering (about 95% from begin_pass, the rest from vk_frame_begin). The command buffers are reused through the implicit reset of vkBeginCommandBuffer (pool flag RESET_COMMAND_BUFFER), and libmali keeps those blocks across that reset.

Change. Before each begin, begin_cb_run resets the command buffer with VK_COMMAND_BUFFER_RESET_RELEASE_RESOURCES_BIT. On by default for VK_DRIVER_ID_ARM_PROPRIETARY only; LARECOMP_VK_CB_RELEASE=0/1 overrides, logged at start like the other per-GPU defaults; a row in docs/vulkan-gpu-defaults.md.

Tested (Vulkan backend only; nothing changes on Metal): Linux, RK3588 (Mali-G610), libmali g29p1, 720p, the Field east of Marin and Tarin's house, 7-minute runs, warm shader cache:

	libmali's block site, 60 s → 420 s	all live malloc, 60 s → 420 s	frame time	GPU time
LARECOMP_VK_CB_RELEASE=0	15 → 57 MB	277 → 259 MB, rising 7 MB/min from 120 s	33.41 ms	16.02 ms
default (on)	9 → 8 MB	180 → 200 MB, flat from 120 s	33.43 ms	16.05 ms

The same change on an earlier build held flat for 7 minutes too. PanVK (Mesa main) and Turnip were already flat and keep the old path.

🤖 Generated with Claude Code

[peiikai]

Thanks, merged into beta_6 in 0c36111. On Arm's driver the command buffers now give their memory back at each begin, so the game no longer grows until it runs out; it goes live with the next release.