Intel Arc (Mesa): ray-traced shadows 3.5-15 FPS in crowded regions vs 25-55 with shadow maps #1104

New issue

Open

opened 2026-10-26 05:43:14 +00:00 by sky_line · 0 comments

sky_line commented

2026-10-26 05:43:14 +00:00

Owner

Ray-traced sun/projector shadows run at 3.5-15 FPS on an Intel Arc B580 (Mesa 26.2.5, Linux) in a crowded region (~3,700 rigged faces); the same scene with shadow maps runs at 25-55 FPS.

Evidence (1.4.2.1 log from the #1059 reporter, 2026-10-26 05:11-05:33 UTC): FPS follows the shadow technique exactly.

Time	Sun shadows	FPS
05:12-05:15	ray traced	7-15
05:15:42-05:16:42	shadow maps (RT over its VRAM budget, 60 s fallback)	49
05:16:42-05:19:28	ray traced again	3.5-5.8
05:19:28-05:20:28	shadow maps (fallback, 60 s)	55
05:20:32 on	shadow maps (fallback 120 / 240 / 480 s)	25-40

The reporter perceived "flicking settings up and down" as the fix; the recoveries actually came from the RT memory-budget fallback (VRAM ~7 GB in use of a 9.6 GB budget; RT working buffers 404-722 MB vs a 491-577 MB headroom budget).

Likely the same class as #1085 (9070 XT / RADV: RT sun visibility far too slow outside NVIDIA). Next:

GPU timers from the reporter: one run with RT shadows on in that region, RenderVKGpuTimers 1, send the log (GPU ms/frame lines name the pass).
Compare with the #1085 findings (merged SUNVIS kernel shared memory + traversal stack, foliage NoOpaque ray, RenderVKRTShadowMerge / RenderVKShadowSplitKernels A/B).
If it is a driver limit, a per-vendor default (tier: Arc -> shadow maps by default, RT opt-in) via #965.

Workaround for users: Preferences -> Sun/moon shadows -> shadow maps.

Ray-traced sun/projector shadows run at 3.5-15 FPS on an Intel Arc B580 (Mesa 26.2.5, Linux) in a crowded region (~3,700 rigged faces); the same scene with shadow maps runs at 25-55 FPS. **Evidence** (1.4.2.1 log from the #1059 reporter, 2026-10-26 05:11-05:33 UTC): FPS follows the shadow technique exactly. | Time | Sun shadows | FPS | |---|---|---| | 05:12-05:15 | ray traced | 7-15 | | 05:15:42-05:16:42 | shadow maps (RT over its VRAM budget, 60 s fallback) | 49 | | 05:16:42-05:19:28 | ray traced again | 3.5-5.8 | | 05:19:28-05:20:28 | shadow maps (fallback, 60 s) | 55 | | 05:20:32 on | shadow maps (fallback 120 / 240 / 480 s) | 25-40 | The reporter perceived "flicking settings up and down" as the fix; the recoveries actually came from the RT memory-budget fallback (VRAM ~7 GB in use of a 9.6 GB budget; RT working buffers 404-722 MB vs a 491-577 MB headroom budget). Likely the same class as #1085 (9070 XT / RADV: RT sun visibility far too slow outside NVIDIA). Next: 1. GPU timers from the reporter: one run with RT shadows on in that region, `RenderVKGpuTimers 1`, send the log (GPU ms/frame lines name the pass). 2. Compare with the #1085 findings (merged SUNVIS kernel shared memory + traversal stack, foliage NoOpaque ray, `RenderVKRTShadowMerge` / `RenderVKShadowSplitKernels` A/B). 3. If it is a driver limit, a per-vendor default (tier: Arc -> shadow maps by default, RT opt-in) via #965. Workaround for users: Preferences -> Sun/moon shadows -> shadow maps.

sky_line added the

labels

2026-10-26 05:43:14 +00:00

sky_line self-assigned this

2026-10-26 05:43:14 +00:00

sky_line added this to the Slipstream project

2026-10-26 05:43:14 +00:00