sky_line/Slipstream

Split the GI trace and RT shadow core kernels into trace and shade passes #1031

New issue

Open

opened 2026-10-25 01:02:05 +00:00 by sky_line · 0 comments

sky_line commented

2026-10-25 01:02:05 +00:00

Owner

Profiled with Nsight Graphics GPU Trace + shader profiler on an RTX 5090 (full GI scene, ~61 FPS, GPU ~16.3 ms/frame). DeferGiTrace (19% of GPU shader samples) and DeferShadowMergeCore / DeferProjStochCore (15%) run at the 255-register cap: 8 warps per SM out of 48, and together cover ~4.5 ms of the frame at 10-15% occupancy. Their machine code is ~0.3-0.9 MB, and the profiler shows instruction-cache stalls (NOINST 53%) inside the deferred kernel, plus a long tail at the end of the GI dispatch.

Split each into a trace pass that writes compact hit records and separate shading pass(es).

Done when: each pass is under 128 registers with no local-memory spill (VKPipeStats), the image matches, and GPU timers at the same scene show a net gain.

Profiled with Nsight Graphics GPU Trace + shader profiler on an RTX 5090 (full GI scene, ~61 FPS, GPU ~16.3 ms/frame). DeferGiTrace (19% of GPU shader samples) and DeferShadowMergeCore / DeferProjStochCore (15%) run at the 255-register cap: 8 warps per SM out of 48, and together cover ~4.5 ms of the frame at 10-15% occupancy. Their machine code is ~0.3-0.9 MB, and the profiler shows instruction-cache stalls (NOINST 53%) inside the deferred kernel, plus a long tail at the end of the GI dispatch. Split each into a trace pass that writes compact hit records and separate shading pass(es). **Done when:** each pass is under 128 registers with no local-memory spill (VKPipeStats), the image matches, and GPU timers at the same scene show a net gain.

sky_line added the

labels

2026-10-25 01:02:05 +00:00

sky_line added this to the Slipstream project

2026-10-25 01:02:05 +00:00

sky_line added the

label

2026-10-25 01:59:11 +00:00

sky_line referenced this issue from a commit

2026-10-25 01:59:12 +00:00

#1031: GI trace split into stages -- the walk, the hit shading, the reuse, the re-shade (RenderVKGISplitKernels)

sky_line referenced this issue from a commit

2026-10-25 01:59:12 +00:00

#1031: generic shadow core split into SELECT + TRACE (RenderVKShadowSplitKernels)

sky_line referenced this issue from a commit

2026-10-25 01:59:12 +00:00

#1031: giDecide clears the split flag before its early returns

sky_line referenced this issue

2026-10-25 02:12:09 +00:00

GI shade and RT reflections: hit/miss compaction and a material/lighting split #1036

sky_line referenced this issue from a commit

2026-10-25 02:44:11 +00:00

#1031: split shadow core TRACE into SUNVIS + VIS (the walks) and SHADE (the terms)

sky_line referenced this issue from a commit

2026-10-25 02:44:11 +00:00

#1031: lean shadow walk -- opaque first, one-kind walk kernels, a slim candidate loop

sky_line referenced this issue from a commit

2026-10-25 02:44:11 +00:00

#1031: RenderVKShadowOpaqueFirst defaults to 0

sky_line referenced this issue

2026-10-25 03:01:22 +00:00

Avatar very noisy for a moment when it goes from occluded to visible #1038

sky_line referenced this issue

2026-10-25 03:19:37 +00:00

[Epic] Shared ray-traced visibility pass (ray queue) for shadows, GI, reflections and lit shaders #1039

sky_line referenced this issue from a commit

2026-10-25 03:44:36 +00:00

#1036: GI shade and RT reflections split -- hit / miss compaction, material / lighting stages

sky_line referenced this issue from a commit

2026-10-25 03:44:36 +00:00

#1031: lazy transform reads in the candidate alpha test and the shadow walk

sky_line referenced this issue from a commit

2026-10-25 03:44:36 +00:00

#1031: dispatch-uniform walk parameters read where used

sky_line referenced this issue from a commit

2026-10-25 03:44:36 +00:00

#1031: isotropic alpha-test LOD for the split shadow walks (RenderVKShadowIsoLod)

sky_line referenced this issue from a commit

2026-10-25 03:44:36 +00:00

#1031: VIS one thread per light sample; a sample's ray inputs parked in shared memory between walks

sky_line referenced this issue from a commit

2026-10-25 04:56:50 +00:00

#1031: fully lit shadow samples read back as fully shadowed (unorm30 overflow)

sky_line self-assigned this

2026-10-26 02:58:25 +00:00

sky_line referenced this issue from a commit

2026-10-26 16:19:10 +00:00

#1039: shadow core queue shaders (RenderVKVisQueue bit 8): SUNVIS / VIS as GEN + WALK

sky_line referenced this issue from a commit

2026-10-26 20:49:37 +00:00

#1039: shadow core queue shaders (RenderVKVisQueue bit 8): SUNVIS / VIS as GEN + WALK

Sign in to join this conversation.

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

sky_line/Slipstream#1031

No description provided.