RT sun shadows: further optimization (shadow-mask pass, reduced-rate tracing, hybrid edges, denoise)

sky_line commented

2026-09-30 03:17:38 +00:00

Owner

Follow-up to #185 (pushed up to 75ca3a315c). Current cost on an RTX 5090, same view: cascades 14.8 ms, ray-traced sun 17.2 ms (+2.4 ms: pass 1 +3.1, pass 2 +0.3, cascades -1.85). Already done: 1 ray/pixel + TAA noise, no rays where N.L <= 0, cascades skipped, transparent receivers opaque-only (blend casters excluded), probe/mirror captures unshadowed.

Candidates (roughly by expected win)

Shadow mask pass instead of in-FS rays. Trace sun visibility in a compute pass right after the depth pre-pass (depth + the thin G-buffer normals already exist for AO/motion) into an R8 mask; the lit FS samples it. Coherent compute rays instead of divergent ray queries inside the heavy lit FS (register pressure / occupancy of every lit pixel pays for the RT code), and it unlocks 2-3 below. Transparent / particle receivers keep per-fragment rays (or sample the mask behind them).
Half-resolution / checkerboard tracing + upsample. Trace 1/2 or 1/4 of the pixels (depth/normal-aware upsample), TAA / temporal accumulation fills the rest. Standard for RT shadows; ~2-4x fewer rays.
Hybrid: trace only near shadow edges (AMD FidelityFX Hybrid Shadows): a cheap low-res cascade (or last frame's mask) classifies tiles as fully lit / fully shadowed; rays only for tiles with an edge. Most of the screen is uniformly lit or in shadow.
Temporal accumulation + denoiser (#85): fewer rays per frame for the same soft penumbra; also removes the reliance on TAA for convergence.
Hardware alpha test: opacity micromaps (#194) -- ~1.3 ms of the cost is the per-leaf candidate alpha test.
Shorter rays / smaller TLAS for the sun: tmax from the shadow range and the RT radius instead of 1024 m; optionally a sun-specific instance selection (light-space extruded view volume) so the sun traces a smaller TLAS; check BLAS build flags (PREFER_FAST_TRACE, compaction) on foliage-heavy scenes.
Cheaper transparent-layer rays: one visibility per pixel for stacked alpha (sample the opaque mask behind + a single ray for the front layer) instead of one ray per layer.
Measurement first: sub-timers that isolate the ray-query cost inside pass 1 (a no-trace variant A/B), plus a per-pixel ray-count / candidate-count debug view to see where the cost concentrates (foliage vs open ground).

Goal

RT sun at or below the cascades' cost (so it can become the default on RT hardware), without visible quality loss (leaf-accurate shadows, stable under TAA).

Related: #185, #194, #85, #84, #157, #164.

Follow-up to #185 (pushed up to 75ca3a315c). Current cost on an RTX 5090, same view: cascades **14.8 ms**, ray-traced sun **17.2 ms** (+2.4 ms: pass 1 +3.1, pass 2 +0.3, cascades -1.85). Already done: 1 ray/pixel + TAA noise, no rays where N.L <= 0, cascades skipped, transparent receivers opaque-only (blend casters excluded), probe/mirror captures unshadowed. ### Candidates (roughly by expected win) 1. **Shadow mask pass instead of in-FS rays.** Trace sun visibility in a compute pass right after the depth pre-pass (depth + the thin G-buffer normals already exist for AO/motion) into an R8 mask; the lit FS samples it. Coherent compute rays instead of divergent ray queries inside the heavy lit FS (register pressure / occupancy of every lit pixel pays for the RT code), and it unlocks 2-3 below. Transparent / particle receivers keep per-fragment rays (or sample the mask behind them). 2. **Half-resolution / checkerboard tracing + upsample.** Trace 1/2 or 1/4 of the pixels (depth/normal-aware upsample), TAA / temporal accumulation fills the rest. Standard for RT shadows; ~2-4x fewer rays. 3. **Hybrid: trace only near shadow edges** (AMD FidelityFX Hybrid Shadows): a cheap low-res cascade (or last frame's mask) classifies tiles as fully lit / fully shadowed; rays only for tiles with an edge. Most of the screen is uniformly lit or in shadow. 4. **Temporal accumulation + denoiser** (#85): fewer rays per frame for the same soft penumbra; also removes the reliance on TAA for convergence. 5. **Hardware alpha test: opacity micromaps** (#194) -- ~1.3 ms of the cost is the per-leaf candidate alpha test. 6. **Shorter rays / smaller TLAS for the sun**: tmax from the shadow range and the RT radius instead of 1024 m; optionally a sun-specific instance selection (light-space extruded view volume) so the sun traces a smaller TLAS; check BLAS build flags (PREFER_FAST_TRACE, compaction) on foliage-heavy scenes. 7. **Cheaper transparent-layer rays**: one visibility per pixel for stacked alpha (sample the opaque mask behind + a single ray for the front layer) instead of one ray per layer. 8. **Measurement first**: sub-timers that isolate the ray-query cost inside pass 1 (a no-trace variant A/B), plus a per-pixel ray-count / candidate-count debug view to see where the cost concentrates (foliage vs open ground). ### Goal RT sun at or below the cascades' cost (so it can become the default on RT hardware), without visible quality loss (leaf-accurate shadows, stable under TAA). Related: #185, #194, #85, #84, #157, #164.

sky_line added the

rt

perf

rendering

labels

2026-09-30 03:17:38 +00:00

sky_line referenced this issue

2026-09-30 03:17:38 +00:00

Ray-traced sun and moon shadows (RenderVKSunShadows 2) #185

sky_line referenced this issue

2026-09-30 06:19:49 +00:00

Ray-traced sun and moon shadows (RenderVKSunShadows 2) #185

sky_line referenced this issue

2026-09-30 06:20:26 +00:00

Ray-traced sun and moon shadows (RenderVKSunShadows 2) #185

sky_line referenced this issue

2026-09-30 20:02:41 +00:00

Perf: Beta 10 slower than Beta 8 on RX 6700 XT (GPU-bound in pass 1; ray-traced sun likely on) #210

sky_line referenced this issue

2026-09-30 21:16:57 +00:00

Translucent coloured shadows: alpha-blend casters tint and attenuate light (stained glass) #212

sky_line referenced this issue

2026-10-01 01:22:47 +00:00

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

sky_line referenced this issue from a commit

2026-10-01 02:11:30 +00:00

VK: ray-traced shadow mask pass (sun + projectors), temporal accumulation

sky_line added this to the Slipstream project

2026-10-06 15:04:15 +00:00

sky_line commented

2026-10-09 04:40:41 +00:00

Author

Owner

Done: the RT shadow-mask pass, reduced-rate tracing and related optimizations shipped. Remaining shadow work is tracked in #567 (shadow-map parity) and #572.

sky_line closed this issue

2026-10-09 04:40:41 +00:00

sky_line referenced this issue from a commit

2026-10-22 08:20:13 +00:00

#195 #85 denoiser: sun / moon visibility channel for the merged RT shadow pass

sky_line referenced this issue from a commit

2026-10-22 08:20:13 +00:00

#195 #85 #869 merged RT shadow pass: sun / moon trace in pass 15, SHADOW reads it

sky_line referenced this issue from a commit

2026-10-22 08:20:13 +00:00

#195 #85 merged RT shadows: host plumbing, settings, shadow-path log

sky_line referenced this issue from a commit