Explore: separate cull scratch buffers for the shadow and mirror culls so the RT-build join can move later (more async overlap) #633

New issue

Open

opened 2026-10-10 09:27:32 +00:00 by sky_line · 0 comments

sky_line commented

2026-10-10 09:27:32 +00:00

Owner

Exploration. Proposed by the 2026-10-10 GPU pipeline audit; not yet implemented.

Idea

The shadow and mirror culls share the cull scratch buffers (cand / cand_meta) with other culls. Because of that, the async RT build has to be joined before the shadow rigged cull. Giving the shadow and mirror culls their own scratch buffers would let the join move to the first TLAS reader in pass 1b, so the RT build overlaps the prepass instead of making graphics wait for it.

Why

About 0.3 ms of q_rt_build likely shows up today as graphics waiting behind it during the prepass (measured 2026-10-10: q_rt_build ~1.3-1.4 ms; q_async_span 5.1 ms against q_async_busy 2.8 ms). It changes no pixels, only overlap.

Cost and risk

Extra VRAM for the duplicate scratch buffers; size them from the existing high-water marks, not the worst case.
Re-check every reader of the TLAS before pass 1b (probe traces, now also on the async queue; shadow RT) and the join barriers.
Validate with synchronization validation on.

Measure

GPU sub ms/frame total= and q_async_span against q_async_busy, in the same sim, before and after.

Related: #256, the roadmap's overlap/async-compute phase.

**Exploration.** Proposed by the 2026-10-10 GPU pipeline audit; not yet implemented. ## Idea The shadow and mirror culls share the cull scratch buffers (`cand` / `cand_meta`) with other culls. Because of that, the async RT build has to be joined before the shadow rigged cull. Giving the shadow and mirror culls their own scratch buffers would let the join move to the **first TLAS reader in pass 1b**, so the RT build overlaps the prepass instead of making graphics wait for it. ## Why About 0.3 ms of `q_rt_build` likely shows up today as graphics waiting behind it during the prepass (measured 2026-10-10: `q_rt_build` ~1.3-1.4 ms; `q_async_span` 5.1 ms against `q_async_busy` 2.8 ms). It changes no pixels, only overlap. ## Cost and risk - Extra VRAM for the duplicate scratch buffers; size them from the existing high-water marks, not the worst case. - Re-check every reader of the TLAS before pass 1b (probe traces, now also on the async queue; shadow RT) and the join barriers. - Validate with synchronization validation on. ## Measure GPU sub ms/frame `total=` and `q_async_span` against `q_async_busy`, in the same sim, before and after. Related: #256, the roadmap's overlap/async-compute phase.

sky_line added the

perf

rendering

labels

2026-10-10 09:27:32 +00:00

sky_line added this to the Slipstream project

2026-10-10 09:27:32 +00:00

sky_line referenced this issue

2026-10-10 11:16:09 +00:00

Epic: visibility-buffer opaque pipeline (explore for FPS) #256