Texture pools: size-class slab pools for 256K-2M textures (fix stranded space of the single TLSF medium pool) #587

New issue

Closed

opened 2026-10-09 03:39:21 +00:00 by sky_line · 1 comment

sky_line commented

2026-10-09 03:39:21 +00:00

Owner

Measured on 2026-10-09 in a heavy sim. Raising the medium texture pool limit from 256 KB to 2 MB had two effects:

Good: live allocations went from 3,254 to 879, and padding from 119 MB to 12 MB. Region-crossing hitches also felt better.
Bad: the single TLSF medium pool (8 MB blocks) sat at 74% occupancy, with 767 MB of free space stranded in partly used blocks. Mixed sizes from 256K to 2M leave holes that later allocations can't use.

Design

Size-class slab pools. SL textures are power-of-two dimensions, so their memory requirements cluster in a series that roughly doubles each step: ~352K, ~704K, ~1.4M, ~2.8M.

Give each class its own VMA pool, block = 8–16 slots, and suballocate exact-size slots.
A freed slot is always reusable by the same class, so there is no external fragmentation in steady state.
Derive the class boundaries from the measured driver sizes. The VKMem tex size bands line now logs the 16 most common exact sizes. Add a few % of headroom per class.
Sizes that don't fit a class go to a small TLSF fallback pool, or stay dedicated above the limit (> ~2–4 MB).
Release empty blocks promptly, and keep per-class block counts bounded.

Stats

Per-class occupancy and stranded bytes go in the VKMem tex pools line.

Backstop (only if still needed after measuring)

A budgeted background defrag: move images into fuller blocks with a GPU copy and a bindless slot rewrite, then free the old image after its fence. Skip images with uploads, sharpen or encode in flight, media textures, and direct image/view holders.

Done when

In a heavy sim:

allocations stay under ~1000;
texture pool occupancy is ≥ 90% steady state;
stranded space is under ~150 MB;
VRAM frees properly after leaving a region.

Measured on 2026-10-09 in a heavy sim. Raising the medium texture pool limit from 256 KB to 2 MB had two effects: - **Good:** live allocations went from 3,254 to 879, and padding from 119 MB to 12 MB. Region-crossing hitches also felt better. - **Bad:** the single TLSF medium pool (8 MB blocks) sat at **74% occupancy, with 767 MB of free space stranded** in partly used blocks. Mixed sizes from 256K to 2M leave holes that later allocations can't use. ## Design **Size-class slab pools.** SL textures are power-of-two dimensions, so their memory requirements cluster in a series that roughly doubles each step: ~352K, ~704K, ~1.4M, ~2.8M. - Give each class its own VMA pool, block = 8–16 slots, and suballocate exact-size slots. - A freed slot is always reusable by the same class, so there is no external fragmentation in steady state. - Derive the class boundaries from the measured driver sizes. The VKMem `tex size bands` line now logs the 16 most common exact sizes. Add a few % of headroom per class. - Sizes that don't fit a class go to a small TLSF fallback pool, or stay dedicated above the limit (> ~2–4 MB). - Release empty blocks promptly, and keep per-class block counts bounded. ## Stats Per-class occupancy and stranded bytes go in the VKMem `tex pools` line. ## Backstop (only if still needed after measuring) A budgeted background defrag: move images into fuller blocks with a GPU copy and a bindless slot rewrite, then free the old image after its fence. Skip images with uploads, sharpen or encode in flight, media textures, and direct image/view holders. ## Done when In a heavy sim: - allocations stay under ~1000; - texture pool occupancy is ≥ 90% steady state; - stranded space is under ~150 MB; - VRAM frees properly after leaving a region.

sky_line added the

labels

2026-10-09 03:39:21 +00:00

sky_line added this to the Slipstream project

2026-10-09 03:39:21 +00:00

sky_line referenced this issue

2026-10-09 04:27:44 +00:00

Texture slab pools: budgeted defrag to release freed-but-not-drained blocks (deferred) #593