RT world path tracer: SER/payload/RIS perf pass + push-constant cleanup

[ComfyFluffy]

Summary

Profiling-driven performance work on the world path tracer (world.rgen.slang), plus two correctness fixes surfaced along the way and one cleanup requested afterward.

Measured, cumulative: ~19.2ms → ~15.7ms/frame (~18%) on the profiled scene.

• Keep the radiance payload out of the SER reorder save set — traceRadianceReordered was reading all 10 payload members back out after TraceRay, holding them live across ReorderThread, then rebuilding them for Invoke. Traversal never writes a radiance payload, so those values were always dead. Rebuilding from trace state instead of carrying the payload across the reorder: ~0.8ms.
• Pack the radiance payload's vector lanes to half3 — albedo/normal/motionPrev/f0 from float3 to half3 (72 → 48 bytes). Payload storage is reserved per trace call site and paid twice per radiance trace, so every byte here is doubled up. hitT stays float (reaches 10000 blocks). No new device feature — Float16/Int16 were already required by packHalf2: ~0.3ms.
• Compact the GPU light record from 48 to 32 bytes — RIS fetches these at random indices, M times per shading vertex; 48 bytes straddled a 64-byte cache line about half the time (2 transactions), 32 bytes never does (1 transaction). Rectangle area is derived (4*|halfU x halfV|) rather than stored — exact, not approximate, since the collector already builds the half-axes that way: ~0.5ms.
• Quarter the RIS candidate count at secondary vertices — probing showed RIS's memory-bound cost was ~5.9ms of an 18ms frame, ~4.3ms of it at secondary vertices (which tolerate far more variance — that light is already integrated over a diffuse lobe before the denoiser sees it) vs ~1.1ms at the primary hit. M/4 past the primary hit, M unchanged at the primary hit: ~1.9ms, visually indistinguishable from a full M/1 test at the same setting.
• Gate RIS emitter NEE past shaded-depth 2 — also drops shadeReservoir's per-vertex shadow ray, which doesn't scale with M and was a large share of what remained: ~0.4ms.

Correctness fixes (found via the above)

• RIS budget was keyed on bounce, not on shaded surfaces. MATERIAL_GLASS/MATERIAL_WATER continue without shading anything, so a wall seen through a pane was bounce 1 but visually the first shaded surface — it could fall off the RIS budget entirely and read as black. Introduced an explicit diffuseDepth counter (incremented only where a surface is actually shaded) and rekeyed the RIS gate and the secondary-vertex divisor on it instead of bounce.
• Same defect in the SSS backlight term (MAX_SSS_BOUNCE → MAX_SSS_DIFFUSE_DEPTH) — leaves/grass seen through glass or water were losing their sun-backlit glow for the same reason. Rekeyed identically.

Cleanup

• Dropped the PBR-toggle flag. worldPush.flags bit 1 gated GGX BRDF vs. a Lambertian fallback, but the Java side packed it as a hardcoded 0b10 — nothing ever cleared it. Removed the flag and every pbr ? x : y site; each now takes the branch that was always live. One site needed care rather than deletion: the particle billboard path passed pbr=false explicitly (independent of the global flag) to suppress a Fresnel rim highlight that F0=0 alone does not zero out at grazing angles. Preserved that behavior by gating on !twoSided instead, since particles are the only twoSided caller.
• Moved every device address in WorldPush to the push-constant block (WorldPushConstants). That struct exists specifically so hit shaders/raygen can read a 64-bit address without dereferencing worldPushAddr first. tableAddr/entityTableAddr in WorldPush were exact duplicates never actually read through worldPush.* anywhere — deleted outright. The five RIS light-buffer addresses were unique and move there for consistency: one struct now holds every device address instead of two. WorldPushConstantsData/WorldPushData are Slang-reflection-generated, so the Java-side change was reordering two constructor call sites to match the new field layout; grows from 40 to 80 bytes, still well under the 128-byte minimum Vulkan guarantee.

Test plan

• ./gradlew build — full build green, including shader compile + spirv-val for both SER (EXT/NV) variants
• ./gradlew test — 34/34 pass
• Manual profiling on the target scene at each step (Nsight live-state + hotspot captures)
• Visual check: indirect lighting quality at SECONDARY_RIS_DIVISOR=4/RIS_MAX_DIFFUSE_DEPTH=2 (kept as committed), particle billboards (no unwanted specular rim), foliage seen through glass/water (SSS backlight restored)
• Reviewer: re-profile on your own scene, since all frame-time numbers above are scene-specific

🤖 Generated with Claude Code