Perf: RT build costs ~3 ms CPU per frame with ray-traced shadows on #93

New issue

Open

opened 2026-09-27 06:56:38 +00:00 by backspace119 · 0 comments

backspace119 commented

2026-09-27 06:56:38 +00:00

Owner

With ray-traced projector shadows on (RenderVKProjectorShadows=2), the CPU-side RT build costs 2.8-3.6 ms/frame (section vk_rtBuild) - more than the GPU work it feeds. The frame was CPU-bound (~14 ms CPU vs ~8 ms GPU, 5090, busy sim).

Look at: per-frame BLAS/TLAS bookkeeping on the CPU (slot scans, build-list assembly, rigged refit lists, descriptor writes - ~21 set writes/frame by design), whether static BLAS state can be cached and only diffs processed, and moving instance/TLAS input generation fully to the existing compute pass. Target: <0.5 ms CPU when nothing changed.

With ray-traced projector shadows on (RenderVKProjectorShadows=2), the CPU-side RT build costs **2.8-3.6 ms/frame** (section `vk_rtBuild`) - more than the GPU work it feeds. The frame was CPU-bound (~14 ms CPU vs ~8 ms GPU, 5090, busy sim). Look at: per-frame BLAS/TLAS bookkeeping on the CPU (slot scans, build-list assembly, rigged refit lists, descriptor writes - ~21 set writes/frame by design), whether static BLAS state can be cached and only diffs processed, and moving instance/TLAS input generation fully to the existing compute pass. Target: <0.5 ms CPU when nothing changed.

backspace119 added the

perf

labels

2026-09-27 06:56:38 +00:00