Shadow maps: near-parity with ray-traced shadows at max projector budget (static caching, caster culling, multiview, filtering)

sky_line commented

2026-10-08 20:11:34 +00:00

Owner

Goal: shadow-map shadows (the path for GPUs without ray tracing, and the fallback) should cost close to the ray-traced path, even at the maximum projector budget.

Baseline (2026-10-08, owner's machine RTX 5090, static scene with 16 projectors, after #564's per-view rigged cull + caster LOD):

	shadow maps	ray traced
GPU frame	14.2 ms	9.7 ms
pass1	9.6 ms	5.1 ms
sun cascades `c_shadows`	2.2 ms	—
projectors `c_projshadow`	2.7–3.5 ms	~0 (in the RT mask)

With the projector budget raised to 32: ~60 FPS on maps vs ~100 on RT.
The remaining cost is mostly static scenery re-rendered into maps, not avatars (the rigged cull costs 0.035 ms).

Target: shadow-map overhead under about 1–1.5 ms over RT at 16 and 32 projectors, without visible quality loss near the camera. Same rules as before: adaptive and targeted, no global cuts, and the user's resolution setting is never changed.

Directions to evaluate (measure each):

Static caching everywhere.
- Projector tiles and sun cascades: render static casters once, then re-render only on light or caster change (object update dirty tracking, per tile/cascade AABB).
- Dynamic casters as an overlay (#564 did this for cached cascades and projector re-composites; extend it so static re-renders are rare, not periodic).
- Cascade scrolling: shift cached cascades in texel steps as the camera moves instead of redrawing.
Caster culling quality.
- GPU per-view culling for static casters too: compute cull into per-view MDI lists, as #564 did for rigged.
- Hi-Z or occlusion for shadow casters: cull casters whose shadow lands on receivers that aren't visible ("shadow receiver culling" — only render casters whose shadow volume intersects the visible receiver bounds).
- Small-caster culling by texel size, per view.
Cheaper rasterisation.
- Multiview / layered rendering: one draw → all cascades or projector tiles (VK_KHR_multiview or layered instancing), cutting draw submission and vertex work.
- Depth-only pipelines with minimal vertex attributes (position-only streams for static casters; meshlet/cluster culling if available).
- Static-mesh shadow LOD (simplified casters for far cascades / small projectors), like #564's rigged caster LOD.
Atlas and resolution policy.
- Projector atlas allocation by screen influence (#564's cov ladder), with time-sliced updates for projectors far from the camera.
- Lower update rate for low-influence projectors.
- Temporal reprojection between updates.
Filtering cost. Check the PCF / soft-shadow sampling cost in pass1 (some of the 4.5 ms delta may be filtering, not map rendering): fewer taps, gather-based PCF, early-out on fully lit or fully shadowed regions.

Done when: the A/B above (and at max projector budget) shows maps within ~1–1.5 ms of RT, with documented quality comparisons.

**Goal:** shadow-map shadows (the path for GPUs without ray tracing, and the fallback) should cost close to the ray-traced path, **even at the maximum projector budget**. **Baseline (2026-10-08, owner's machine RTX 5090, static scene with 16 projectors, after #564's per-view rigged cull + caster LOD):** | | shadow maps | ray traced | |---|---|---| | GPU frame | 14.2 ms | 9.7 ms | | pass1 | 9.6 ms | 5.1 ms | | sun cascades `c_shadows` | 2.2 ms | — | | projectors `c_projshadow` | 2.7–3.5 ms | ~0 (in the RT mask) | - With the projector budget raised to 32: ~60 FPS on maps vs ~100 on RT. - The remaining cost is mostly **static scenery** re-rendered into maps, not avatars (the rigged cull costs 0.035 ms). **Target:** shadow-map overhead under about 1–1.5 ms over RT at 16 and 32 projectors, without visible quality loss near the camera. Same rules as before: adaptive and targeted, no global cuts, and the user's resolution setting is never changed. **Directions to evaluate (measure each):** 1. **Static caching everywhere.** - Projector tiles and sun cascades: render static casters once, then re-render only on light or caster change (object update dirty tracking, per tile/cascade AABB). - Dynamic casters as an overlay (#564 did this for cached cascades and projector re-composites; extend it so static re-renders are rare, not periodic). - Cascade scrolling: shift cached cascades in texel steps as the camera moves instead of redrawing. 2. **Caster culling quality.** - GPU per-view culling for **static** casters too: compute cull into per-view MDI lists, as #564 did for rigged. - Hi-Z or occlusion for shadow casters: cull casters whose shadow lands on receivers that aren't visible ("shadow receiver culling" — only render casters whose shadow volume intersects the visible receiver bounds). - Small-caster culling by texel size, per view. 3. **Cheaper rasterisation.** - Multiview / layered rendering: one draw → all cascades or projector tiles (VK_KHR_multiview or layered instancing), cutting draw submission and vertex work. - Depth-only pipelines with minimal vertex attributes (position-only streams for static casters; meshlet/cluster culling if available). - Static-mesh shadow LOD (simplified casters for far cascades / small projectors), like #564's rigged caster LOD. 4. **Atlas and resolution policy.** - Projector atlas allocation by screen influence (#564's cov ladder), with time-sliced updates for projectors far from the camera. - Lower update rate for low-influence projectors. - Temporal reprojection between updates. 5. **Filtering cost.** Check the PCF / soft-shadow sampling cost in pass1 (some of the 4.5 ms delta may be filtering, not map rendering): fewer taps, gather-based PCF, early-out on fully lit or fully shadowed regions. **Done when:** the A/B above (and at max projector budget) shows maps within ~1–1.5 ms of RT, with documented quality comparisons.

sky_line added the

labels

2026-10-08 20:11:34 +00:00

sky_line added this to the Slipstream project

2026-10-08 20:11:34 +00:00

sky_line referenced this issue

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

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

sky_line referenced this issue

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

RT sun shadows: further optimization (shadow-mask pass, reduced-rate tracing, hybrid edges, denoise) #195

sky_line referenced this issue

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

Some projector light still bleeds through walls (shadow maps) #3

sky_line commented

2026-10-10 11:31:58 +00:00

Author

Owner

Started 2026-10-10. Fresh owner measurement: RTX 5090, 4K with FSR, busy sim, 32 projectors shadowed.

	FPS	GPU frame	sun cascades	projectors
all ray traced	~90	10.8 ms	—	— (in c_rtmask 0.82 ms)
sun/moon on maps, projectors RT	~66	14.8 ms	2.55 ms	—
sun/moon + projectors on maps	~44	22.5 ms	2.0–2.1 ms	9.4–9.8 ms

Maps cost ~+4 ms for sun/moon and ~+12 ms with 32 projectors, over RT.

Order of work:

Projector tiles: instrument refresh causes, make static re-renders event-driven, GPU per-tile static caster cull, batch dirty tiles.
Sun cascades: texel-step scrolling, per-cascade static cull, dirty-only static re-render.
Filtering cost: deferred light 1.13 → ~1.4 ms with maps.
Mode switch: check whether both paths run during a shadow-mode change (RT mask/AO doubled for ~20 s during the switch).

Started 2026-10-10. Fresh owner measurement: RTX 5090, 4K with FSR, busy sim, 32 projectors shadowed. | | FPS | GPU frame | sun cascades | projectors | |---|---|---|---|---| | all ray traced | ~90 | 10.8 ms | — | — (in c_rtmask 0.82 ms) | | sun/moon on maps, projectors RT | ~66 | 14.8 ms | 2.55 ms | — | | sun/moon + projectors on maps | ~44 | 22.5 ms | 2.0–2.1 ms | **9.4–9.8 ms** | Maps cost ~+4 ms for sun/moon and ~+12 ms with 32 projectors, over RT. Order of work: 1. **Projector tiles:** instrument refresh causes, make static re-renders event-driven, GPU per-tile static caster cull, batch dirty tiles. 2. **Sun cascades:** texel-step scrolling, per-cascade static cull, dirty-only static re-render. 3. **Filtering cost:** deferred light 1.13 → ~1.4 ms with maps. 4. **Mode switch:** check whether both paths run during a shadow-mode change (RT mask/AO doubled for ~20 s during the switch).

sky_line added

and removed

labels

2026-10-10 11:31:58 +00:00

sky_line referenced this issue

2026-10-10 11:55:54 +00:00

Point light shadows (ray traced + cube shadow-map fallback) #439

sky_line commented

2026-10-10 12:18:41 +00:00

Author

Owner

Correction 2026-10-10: the baseline table above mis-attributed the sun cost

Re-reading the log (11:26-11:29), the "all ray traced ~90 fps / 10.8 ms" row had no sun shadow at all: the sun switched on at 11:27:32.

window	state	GPU ms
≤11:27:30	no sun shadow, projectors RT	10.8
11:27:40-55	sun RT + projectors RT	13.8-14.05
11:28:05-25	sun on maps + projectors RT	14.7-14.8 (+0.85 ms over RT sun)
11:28:35+	everything on maps (32 projectors)	22.5-24.0 (+9.3 ms = projectors)

Sun shadow maps are already within the target.
Projector tiles are the whole problem. The tile cache missed 100%: projTilesRendered=32 projTilesCached=0 projDirtLight=12 projDirtChanged=20 every frame.
The "mode-switch double-run" was not real. That window was an RT-sun period.
c_shadows has a ~1.9 ms floor even with no cascades (likely a cross-queue wait), so it isn't a reliable measure of cascade cost.

Work merged:

counters;
pose-refresh and sub-texel light-move false-invalidation fixes;
a static/dynamic caster split for projector tiles (movers drawn as an overlay, baked in after 2 s);
invalidation from the instance store (also fixes new rezzes appearing late in cached tiles and deleted objects' shadows staying);
a moving-light budget;
a sub-texel caster cull fix;
event-driven cached sun cascades (modes 1/2).

Expected projector steady state ~3-4 ms; the next step depends on c_proj_cull vs g_proj_static in the new counters.

## Correction 2026-10-10: the baseline table above mis-attributed the sun cost Re-reading the log (11:26-11:29), the "all ray traced ~90 fps / 10.8 ms" row had **no sun shadow at all**: the sun switched on at 11:27:32. | window | state | GPU ms | |---|---|---| | ≤11:27:30 | no sun shadow, projectors RT | 10.8 | | 11:27:40-55 | sun RT + projectors RT | 13.8-14.05 | | 11:28:05-25 | sun on maps + projectors RT | 14.7-14.8 (**+0.85 ms** over RT sun) | | 11:28:35+ | everything on maps (32 projectors) | 22.5-24.0 (**+9.3 ms = projectors**) | - **Sun shadow maps are already within the target.** - **Projector tiles are the whole problem.** The tile cache missed 100%: `projTilesRendered=32 projTilesCached=0 projDirtLight=12 projDirtChanged=20` every frame. - **The "mode-switch double-run" was not real.** That window was an RT-sun period. - **`c_shadows` has a ~1.9 ms floor** even with no cascades (likely a cross-queue wait), so it isn't a reliable measure of cascade cost. **Work merged:** - counters; - pose-refresh and sub-texel light-move false-invalidation fixes; - a static/dynamic caster split for projector tiles (movers drawn as an overlay, baked in after 2 s); - invalidation from the instance store (also fixes new rezzes appearing late in cached tiles and deleted objects' shadows staying); - a moving-light budget; - a sub-texel caster cull fix; - event-driven cached sun cascades (modes 1/2). Expected projector steady state ~3-4 ms; the next step depends on `c_proj_cull` vs `g_proj_static` in the new counters.

sky_line referenced this issue from a commit

2026-10-10 22:19:25 +00:00

#567 projector shadows: invalidation counters + two false-invalidation fixes

sky_line referenced this issue from a commit

2026-10-10 22:19:25 +00:00

#567 projector shadows: static / dynamic caster split (movers, spinners, attachments -> overlay)

sky_line referenced this issue from a commit

2026-10-10 22:19:25 +00:00

#567 regen vk_world_shaders_spv.h (clean_cull.comp static / dynamic caster split)

sky_line referenced this issue from a commit

2026-10-10 22:19:25 +00:00

#567 projector shadows: invalidate from the instance store (shadow-relevant changes only)

sky_line referenced this issue from a commit

2026-10-10 22:19:25 +00:00

#567 shadows: moving-light budget under load; sub-texel caster cull (fix + projector tiles)

sky_line referenced this issue from a commit

2026-10-10 22:19:25 +00:00

#567 regen vk_world_shaders_spv.h (casterOk unpadded extent + projector sub-texel cull)

sky_line referenced this issue from a commit

2026-10-10 22:19:25 +00:00

#567 registration spheres: coalesce an object's faces, cap 8192

sky_line referenced this issue from a commit

2026-10-10 22:19:25 +00:00

#567 sun cascades: static / dynamic split + event-driven cache in the cached modes

sky_line referenced this issue from a commit

2026-10-10 22:19:25 +00:00

#567 regen vk_world_shaders_spv.h (cascade overlay cull: u_cascade 5)

sky_line referenced this issue from a commit

2026-10-10 22:19:25 +00:00

#567 shadowCascadesDrawn: count cascades actually drawn (after the RT-sun skip)

sky_line referenced this issue from a commit

2026-10-10 22:19:25 +00:00

Merge #567 shadow-map caching; regenerate SPIR-V

sky_line referenced this issue

2026-10-11 10:06:39 +00:00

Sun shadows bounce back and forth with a static object and static sun (all sun shadows) #674

sky_line added

and removed

labels

2026-10-11 16:37:24 +00:00

sky_line commented

2026-10-11 16:37:25 +00:00

Author

Owner

Shipped in 1.3.0 Beta 2: projector static/dynamic caster split, event-driven tile and cascade caches, invalidation from the instance store, sub-texel caster cull. Sun cascades are already within target. Still to do: re-measure projector tiles at 16 and 32 projectors (expected ~3-4 ms), then GPU per-tile static caster cull and the filtering cost if the numbers still miss the 1-1.5 ms target.

sky_line referenced this issue from a commit

2026-10-14 05:50:50 +00:00

Cluster LOD: draw the cluster path in shadows, mirror, probe faces and the GPU pick

sky_line commented

2026-10-15 01:09:38 +00:00

Author

Owner

Follow-up item (2026-10-14 status audit): the event-driven cascade cache (RenderVKShadowEventCache + static/dynamic split + instance-store invalidation) only runs with RenderVKShadowUpdate 1/2; the default is 0 (Realtime: all 4 cascades every frame, static layers freed). Evaluate making the event-driven cache the default (or auto-selected) for shadow-map users, after an A/B of quality (texel-step swimming, light-move threshold) and GPU ms on a static and a busy scene.

Follow-up item (2026-10-14 status audit): the event-driven cascade cache (RenderVKShadowEventCache + static/dynamic split + instance-store invalidation) only runs with `RenderVKShadowUpdate` 1/2; the default is 0 (Realtime: all 4 cascades every frame, static layers freed). Evaluate making the event-driven cache the default (or auto-selected) for shadow-map users, after an A/B of quality (texel-step swimming, light-move threshold) and GPU ms on a static and a busy scene.

sky_line referenced this issue from a commit

2026-10-20 13:44:59 +00:00

#783: M3 renderer: the snow surface mesh in the deferred G-buffer and the sun cascades

sky_line referenced this issue from a commit

2026-10-22 19:32:55 +00:00

#567 projector shadows: per-frame budget for soft static re-renders

sky_line referenced this issue from a commit

2026-10-22 19:32:55 +00:00

#567 shadow view API: projector atlas can host cube faces (additive)

sky_line referenced this issue from a commit

2026-10-22 19:32:55 +00:00

#567 projector shadows: #825 cascade hold defers through the static budget