Uploads decode on a background dispatcher and the submit carries a
timeline semaphore wait, so the GPU waits for the texture rather than
the render thread. The gather of the application's compressed bytes
moves there too when they are already mapped; unmapping or freeing that
memory drains the decodes still reading it.
Staged copies gain an explicit host-write barrier. A queue submission
only makes host writes from before it visible to the device, and these
writes land after the submit.
Devices without timelineSemaphore, and submits whose pNext already sizes
arrays by their semaphore counts, decode inline as before.
Blade Ball, median of three: worst vkQueueSubmit 140ms -> 1ms, p99
4.9ms -> 0.05ms, total 2534ms -> 309ms. Frame interval unchanged; its
tail is the engine's own main-thread step.
Release the GameMode request when the daemon narrows the CPU mask; the pin
concentrates Roblox's worker pool and is re-applied every few seconds, so
ending the request is the only thing that lifts it. `gamemode: on` keeps it.
Emit texture uploads on worker threads instead of serially inside
vkQueueSubmit, in two passes so mip 0 lands before the levels scaled from it.
Decode ETC2 through a persistent pool and lower the parallel floor to one
128x128 image, so small batches stop running single-threaded.
Document all three in docs/PERFORMANCE.md.
vii