Allocation failures: GPU memory requests retry, then drop -- never crash

sky_line commented

2026-10-12 06:11:42 +00:00

Owner

Policy (owner, 2026-10-12): a failed allocation is never a crash. Default: keep what we have, retry the request a few frames later (space may free up), then drop that asset/request and log the failure (rate-limited, with what failed, size, and the current RAM/VRAM pressure state). Better a missing texture/mesh than a dead viewer -- and when someone says "things aren't loading", the log tells us it's memory pressure.

Scope (Vulkan / video memory)

RT: wrap acceleration-structure builds and TLAS/aux buffers in a soft-allocation scope so the existing fallbacks (skip build, TLAS fallback, abort frame) actually run.
Upload staging (stageFor) soft by default; large temporary host-visible buffers get a pre-flight too.
Growth paths keep their old capacity on failure and retry later: instance tables (ensureCap last resort), occlusion scratch, particle tables, rig cache, alpha-sort keys.
Render targets rebuilt by the VRAM ladder (scale targets, mirrors, probes, temporal) fail softly: keep the previous targets or drop the feature, then step the ladder.
Out-of-memory from command-buffer end, queue submit, image-view and descriptor allocation: drop the frame, step the ladder, log -- not fatal.
The soft-allocation flag is thread-safe and restored (not reset) after scoped use.
Verify with the fault-injection switches (SLVkInjectAllocFail, hard mode) on a teleport loop until no fatal site remains.

**Policy (owner, 2026-10-12):** a failed allocation is never a crash. Default: keep what we have, retry the request a few frames later (space may free up), then drop that asset/request and log the failure (rate-limited, with what failed, size, and the current RAM/VRAM pressure state). Better a missing texture/mesh than a dead viewer -- and when someone says "things aren't loading", the log tells us it's memory pressure. **Scope (Vulkan / video memory)** - RT: wrap acceleration-structure builds and TLAS/aux buffers in a soft-allocation scope so the existing fallbacks (skip build, TLAS fallback, abort frame) actually run. - Upload staging (`stageFor`) soft by default; large temporary host-visible buffers get a pre-flight too. - Growth paths keep their old capacity on failure and retry later: instance tables (`ensureCap` last resort), occlusion scratch, particle tables, rig cache, alpha-sort keys. - Render targets rebuilt by the VRAM ladder (scale targets, mirrors, probes, temporal) fail softly: keep the previous targets or drop the feature, then step the ladder. - Out-of-memory from command-buffer end, queue submit, image-view and descriptor allocation: drop the frame, step the ladder, log -- not fatal. - The soft-allocation flag is thread-safe and restored (not reset) after scoped use. - Verify with the fault-injection switches (`SLVkInjectAllocFail`, hard mode) on a teleport loop until no fatal site remains.

sky_line added the

status

in-progress

label

2026-10-12 06:11:42 +00:00

sky_line added this to the Slipstream project

2026-10-12 06:11:42 +00:00

sky_line referenced this issue from a commit

2026-10-12 12:33:14 +00:00

#697: per-thread soft-allocation mode; texture pass restores the previous mode

sky_line referenced this issue from a commit

2026-10-12 12:33:14 +00:00

#697: RT acceleration-structure build allocates soft

sky_line referenced this issue from a commit

2026-10-12 12:33:14 +00:00

#699: GPU wait accounting per call site + redundant device-idle elision; #697 low-memory counters

sky_line referenced this issue from a commit

2026-10-12 12:33:14 +00:00

#697: upload staging is soft; refused streams retry next frame

sky_line referenced this issue from a commit

2026-10-12 12:33:14 +00:00

#697: instance-id buffers, cull scratch, probe lists and texel demand grow softly

sky_line referenced this issue from a commit

2026-10-12 12:33:14 +00:00

#697: occlusion / Hi-Z, particle emitters, skin cache, alpha-sort keys, anim buffers grow softly

sky_line referenced this issue from a commit

2026-10-12 12:33:14 +00:00

#697: ladder-driven render targets fail softly (scaled targets, mirror)

sky_line referenced this issue from a commit

2026-10-12 12:33:14 +00:00

#697: out-of-memory at frame end / submit drops the frame; soft image views, binds and lazy descriptor sets

sky_line referenced this issue from a commit

2026-10-12 12:33:14 +00:00

#697: textures whose uploads keep failing are dropped for a cooldown, then retried

sky_line referenced this issue from a commit

2026-10-12 12:33:14 +00:00

#697: clear the frame-OOM flag at the start of every async frame

sky_line referenced this issue from a commit

2026-10-12 16:50:36 +00:00

#692: RT rigged groups never stay out of the TLAS silently

sky_line referenced this issue from a commit

2026-10-14 05:50:50 +00:00

Cluster LOD 2c (#749): tolerance controller, per-view error, pre-pass, core fixes

sky_line referenced this issue from a commit

2026-10-14 05:50:50 +00:00

RT: per-mesh cluster path rule; headroom reserve under the VRAM ladder

sky_line referenced this issue

2026-10-16 13:28:03 +00:00

Deferred lighting: move the remaining uber-kernel cases to the split path (debug views, snapshots, allocation failure) #799

sky_line referenced this issue from a commit

2026-10-16 17:53:48 +00:00

VK: split deferred lighting is the default; remove the auto A/B mode

sky_line referenced this issue from a commit

2026-10-17 23:28:39 +00:00

VK: per-bucket max draw counts (Pascal pays for maxDrawCount: -15 ms at 4K on a GTX 1080)

sky_line commented

2026-10-18 22:26:36 +00:00

Author

Owner

Fixed in Slipstream 1.3.0 (stable): https://git.slipstreamviewer.org/sky_line/Slipstream/releases/tag/v1.3.0

If you still see this on 1.3.0, reply here and we'll reopen it.

Fixed in **Slipstream 1.3.0** (stable): https://git.slipstreamviewer.org/sky_line/Slipstream/releases/tag/v1.3.0 If you still see this on 1.3.0, reply here and we'll reopen it.

sky_line closed this issue

2026-10-18 22:26:36 +00:00

sky_line referenced this issue from a commit

2026-10-19 11:20:27 +00:00

RT reflections (#859): renderer -- mark target, RT reflect pass, consumer + stand-down, settings

sky_line referenced this issue from a commit

2026-10-19 11:20:27 +00:00

RT reflections (#859) round 2: renderer -- tile list, trace/resolve/forward/uber kernels, targets, settings

sky_line referenced this issue from a commit

2026-10-21 02:45:33 +00:00

#85 phase 2: diffuse / specular lighting denoiser (LLCleanDenoiserVK)

sky_line referenced this issue from a commit