The directional-shadow topology rebuilt on every streaming-churn frame
and was the measured body of the run-hitch stalls (701 of 708 baseline
stalls alloc-correlated):
- The draw sort comparer's enum-vs-enum CompareTo bound to
Enum.CompareTo(object) and boxed BOTH operands on every comparison —
a constant ~38.9 MB of garbage per topology rebuild (~4M boxes),
handing the GC a forced gen0 collection mid-frame. The full ~100k-draw
sort is replaced outright: draws hash-group by exact DrawKey in one
O(n) pass over retained chained-index arrays, and only the
few-thousand DISTINCT group keys sort (order-preserving packed
material|cull|firstIndex|baseVertex + count|slot|layer|foliage keys,
first-appearance tie-break) — bit-identical emission order to the old
stable sort, near-zero allocation, and no per-draw comparisons at all.
- The caster frame sorts 4-byte indices keyed on SortKey.Value instead
of shuffling multi-hundred-byte records through a boxing comparer.
- Owner-approved pipelining: on a frame whose shadow inputs just changed
(the same frame already paying frame-view/landscape rebuilds), the
caster-frame and prepared-draws topology rebuilds defer to the next
quieter frame, capped at two consecutive deferrals — inside the GPU
fence depth, so retained draws never reference a released arena range.
First build, generation change, caster BuildSequence change, and
journal overflow force the immediate path; deferred refreshes skip
identity-mismatched journal rows.
Owner-accepted in both presentation modes: stall frames 5.8/s -> ~0.45/s
uncapped (0.49/s capped), median stall 20.3 -> 13.7 ms, >25 ms frames
near zero, 275 fps uncapped baseline restored. Allocation gate: a warmed
topology rebuild must allocate <2 KiB (DirectionalShadowPreparedDrawTests).
docs/ISSUES.md carries the full evidence trail; the residual
content-proportional rebuild milliseconds are filed as the
incremental-topology successor, and the pre-existing town-view scaling
latch is filed as #432.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Opus dual-lens review of the three VM6 commits (0930c35d, 39e8408c,
6cc5e183) found two blockers and two should-fix issues; all landed here
along with the review's nits and documentation corrections.
Blockers:
- A1: the procedural-scenery classifier tested bit 31 alone instead of
the full top nibble (0xF000_0000 == 0x8000_0000), so it also matched
LandblockStaticEntityIdAllocator's 0xC... namespace (fences/gates/
building shells with a cutout subset), the 0xDA11_D0xx paperdoll id,
and the 0xFFFF_FF01 portal-tunnel id as procedural scenery — all
three would have swayed. ProceduralSceneryIdAllocator.IsInNamespace
now does the exact top-nibble test; FoliageWindClassification
delegates to it.
- A2: GroupKey (the receiver's instance-batching key) did not carry
FoliageFlags while the caster's dedup key already did, so a scenery
instance and a non-scenery instance sharing a mesh subset coalesced
into one receiver InstanceGroup whose flags were last-writer-wins —
disagreeing with the correctly-keyed caster. GroupKey now carries
FoliageFlags, computed before key construction and set exactly once
at group creation; the imperative re-stamp is gone, and CachedBatch's
now-redundant FoliageFlags field is removed.
Should-fix:
- A3: the world receiver pass bound UniformAtmosphericFrame only by
accident (leftover from the caster pass, which runs first each
frame, since Vulkan binding state isn't reset between passes).
DirectionalShadowFrameBinding now carries the caster's exact
AtmosphericFrameBufferBinding and BindDirectionalShadowReceiver binds
it explicitly.
- A4: a Setup-composed tree's opaque trunk part never got the trunk
flag because HasCutoutSubset is cached per GfxObj part, not per
entity. FoliageWindClassification.ComputeEntityHasCutoutSubset now
ORs HasCutoutSubset across an entity's resolved sibling parts once
per entity, threaded into ClassifyBatches/AddDirectionalShadowBatches
via a new optional override parameter.
Nits: A5 hashes the per-vertex flutter seed relative to the instance
origin instead of absolute world XY (fp32 sin() precision loss at far
landblock corners), mirrored in both foliage_wind.glsl and
FoliageWindModel; A7 documents the max(maxHeight, 0.5) divide-guard as
a deliberate pseudocode divergence; A8 switches FoliageWindExclusions'
construction to ToFrozenSet() and softens the "never stale" doc
comment to "no slower than one frame behind."
Tests added: top-nibble classification (0xFFFFFFFFu now correctly
false), GroupKey inequality across entity-driven scenery/landblock-
static classification, a caster-batch test proving the same pairing
never coalesces, ComputeEntityHasCutoutSubset unit + end-to-end
two-part-Setup tests, the caster→receiver AtmosphericFrame binding
carry-through, flutter-hash translation invariance relative to
instance origin, and a Storm-wind mid-height displacement floor
guarding against a "no motion" regression.
Docs: plan VM6 body corrected to the five-row WeatherKind table, "bits
1 and 2", "all four" caster shaders, and top-nibble wording throughout;
the owner gate checklist's Rain/Storm step; the stale v1-only shader-
interface compatibility entry; semantic-bindings-v1.md's v2 members
folded into the main 192-byte block; the IA-25 register row's top-
nibble wording; AtmosphericFrameInputs.cs's ABI size reference.
foliage_wind.glsl's A5 change recompiled exactly the five shaders that
include it (mesh_atmospheric.vert, the four directional_shadow_world_*
casters) plus the manifest; no other .spv changed.
Verify: Release build 0 warnings/0 errors. App hermetic-lane filter
6,041/0 failed (no environment-specific failures this run).
RenderPackValidator 30/30. Full hermetic-filtered solution: 15,269/0
failed across 15 projects.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
