Perf: geometry cost — depth prepass + pass1 ~21 ms on a GTX 1080 Ti #123

New issue

Open

opened 2026-09-27 18:33:24 +00:00 by backspace119 · 0 comments

backspace119 commented

2026-09-27 18:33:24 +00:00

Owner

GTX 1080 Ti @ 1918×1008, RenderVKGpuTimers (see #98): with shadows off, g_prepass is 6.8 ms and g_pass1 14.2 ms. The cascades then re-draw the scene 4×, for another 33–38 ms at 2048 realtime. That suggests we're vertex/primitive bound: too many triangles for mid-range GPUs.

Ideas:

LOD bias per preset: for mesh/prim LOD (RenderVolumeLODFactor, avatar LOD) and in the VK LOD selection.
Shadow casters:
- use a lower LOD in the shadow passes;
- cull casters by their projected size per cascade (small objects contribute nothing in far cascades);
- optionally skip far cascades for tiny/alpha casters.
Depth prepass: make it optional on low tiers, or restrict it to large/occluding geometry.
Measure first: triangle/instance counts per pass on the same scene. Add counters to the VKGpu log.

Part of #98.

GTX 1080 Ti @ 1918×1008, `RenderVKGpuTimers` (see #98): with shadows off, `g_prepass` is 6.8 ms and `g_pass1` 14.2 ms. The cascades then re-draw the scene 4×, for another 33–38 ms at 2048 realtime. That suggests we're vertex/primitive bound: too many triangles for mid-range GPUs. Ideas: - **LOD bias per preset:** for mesh/prim LOD (RenderVolumeLODFactor, avatar LOD) and in the VK LOD selection. - **Shadow casters:** - use a lower LOD in the shadow passes; - cull casters by their projected size per cascade (small objects contribute nothing in far cascades); - optionally skip far cascades for tiny/alpha casters. - **Depth prepass:** make it optional on low tiers, or restrict it to large/occluding geometry. - **Measure first:** triangle/instance counts per pass on the same scene. Add counters to the VKGpu log. Part of #98.