docs: spec CODEX-TASK-002 (column-memo thrash) and aim the PERF work

Three orchestration findings on the parked PERF item, none measured yet:

1. The perf A/B needs no new code -- two Insights traces on a deterministic
   world are a clean before/after. But tile skipping changes how many times
   GenerateMesh runs, so totals conflate "cheaper per tile" with "fewer
   tiles". TilesMeshed from task 001 is the denominator. 001 is therefore a
   PREREQUISITE for the perf work, not a parallel item as the handoff had it.

2. The suspects don't share an archetype: SurfaceWorld is the column memo,
   TunnelNetwork is 19 virtual calls per voxel (16 + 3 structural post).
   One lumped number can't be acted on -- measure one strate at a time.

3. GetColumn is a direct-mapped hashed table where GSurfColCache is a
   direct-indexed box, and the mesher pre-samples Z-OUTERMOST (verified,
   VoxelMarchingCubesMesher.cpp ~226) so every column is revisited ~34x per
   tile. The table's sizing comment reasons about fullness; a direct-mapped
   table evicts on collision. At load factor 0.28 that is still ~25% of
   columns recomputing the full height stack every Z plane -- derived ~9x.

Task 002 measures that and fixes nothing, so the instrument and the fix
cannot land in the same build and make each other unreadable.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
2026-08-16 15:45:44 +02:00
parent 57002b35cf
commit 7313201e91
2 changed files with 167 additions and 0 deletions
+112
View File
@@ -0,0 +1,112 @@
# Codex task 002 — is the op-stack column memo thrashing? Count, don't guess.
**Owner:** Codex (Model Luna, xHigh) · **Orchestrator:** Claude · **Branch:** `experimental`
**Status:** specified, not started
**Depends on:** `CODEX-TASK-001` — this adds counters to the **same** `stat VoxelForge` group.
Do 001 first; this task assumes `VoxelStats.h` already exists.
---
## Why this exists
The operator-stack density path is measurably slower than the `switch` it replaces, and the cause
has never been attributed. There are three standing suspects. **This task measures the first one,
and does not fix anything.** That order is deliberate: `AUDIT §C10` cost six builds and five refuted
hypotheses by reasoning first, and the last session re-learned it.
### The hypothesis, derived from the code
`FSurfaceColumnSource::GetColumn` (`VoxelDensityOpStack.cpp` ~615) memoises a computed column in a
**direct-mapped, 4096-entry `thread_local` table**, indexed by a hash of the two XY floats:
```cpp
struct FSlot { uint64 Key; float X, Y; FColumn C; }; // 40 bytes
thread_local FSlot Slots[4096] = {}; // 160 KB per worker
const uint32 Idx = ((HX * 0x9E3779B9u) ^ (HY * 0x85EBCA6Bu)) >> 20; // [0,4095]
```
The original path, `GSurfColCache` in `GetDensityAt`, is instead a **direct-indexed box**:
`CI = (IY - Box.BaseY) * Dim + (IX - Box.BaseX)`, with a `Computed[CI]` flag. **No hash, therefore
no collisions, therefore every column is computed exactly once.**
Now the sampling order, which is the premise that makes this bite. `FVoxelMarchingCubesMesher`
pre-samples with **Z as the OUTERMOST loop** (`VoxelMarchingCubesMesher.cpp` ~226):
```cpp
for (int32 gz = GzLo; gz <= GzHi; gz++)
for (int32 gy = -1; gy <= GridDim; gy++)
for (int32 gx = -1; gx <= GridDim; gx++)
Generator->GetDensityAt(...);
```
So the mesher sweeps a **whole XY plane at every Z level**. Every column in the tile is revisited
once per Z plane — roughly 34 times.
The table's own comment sized it for this: *"A chunk is CHUNK_SIZE² columns (1024), so the first
draft's 256 entries could not even hold one chunk and thrashed inside a single tile. 4096 covers
four chunks."* **That reasoning has a gap.** A direct-mapped table does not need to be full to
evict — it needs two live keys to collide. At ~1156 columns per plane in 4096 slots (load factor
0.28), the expected number of columns sharing a slot with another is **~285, about 25 %**. Those
columns evict each other, miss again on the next Z plane, and recompute the **entire height stack**
structural source, cliff (four structural resamples), terrace, layer-line, beach, and the ceiling
stack — every single plane.
Order of magnitude if that is right: ~1156 column computations on the original path versus
~1156 + 285 × 34 ≈ **10 000** on the op path. Roughly **9×** the column work, on the plugin's most
expensive archetype.
**That is a derivation, not a measurement, and it is exactly the kind of confident chain this
project has watched reverse six times.** Hence: count first.
## What to build — two counters, no behaviour change
Add to the existing `stat VoxelForge` group from task 001:
| counter | incremented when |
|---|---|
| `ColumnMemoHit` | `GetColumn` found a live entry (the `if` body did **not** run) |
| `ColumnMemoMiss` | `GetColumn` recomputed (the `if` body ran) |
That is the entire change. Two `INC_DWORD_STAT` calls inside `FSurfaceColumnSource::GetColumn`,
around the existing `if (S.Key != ColumnKey || S.X != WorldX || S.Y != WorldY)`.
## ⚠️ Invariants
1. **Change nothing else in `GetColumn`.** Not the table size, not the hash, not the key comparison.
The point of this task is to produce a number that decides whether the fix is worth writing; a
change bundled in makes the number unattributable. **Do not "obviously improve" the table while
you are in there** — if the fix ships in the same build as the instrument, we learn nothing, and
this project has a written lesson about exactly that.
2. **`GetColumn` is `const` and runs on mesher workers.** `INC_DWORD_STAT`, never a `static int32++`
— same rule as task 001, same reason.
3. **The key comparison is load-bearing and must stay complete.** `S.Key != ColumnKey || S.X != X ||
S.Y != Y` — the full key is compared on every touch precisely so a hash collision can only cost a
recompute and never return **someone else's column**. Do not shorten it to feed a counter.
4. **Zero cost when `STATS == 0`.** No value computed outside the macros.
5. `FCaveCliffMod` and the overhang read this same memo through
`Column->GetColumn(...)` (~line 864). They are legitimate traffic and must be counted, not
excluded — they are part of why a miss is expensive.
## Acceptance
Fly a route through a **SurfaceWorld** strate that has `bUseOperatorStack` ticked, `stat VoxelForge`
on screen, and read the ratio.
- **`ColumnMemoMiss` ≈ one per distinct column per tile** (hit rate climbing toward ~97 % as Z
planes accumulate) ⇒ **the hypothesis is WRONG**, the table is fine, and the perf cost is
elsewhere. Record that and move to suspect 2 (per-voxel virtual dispatch). *A negative result here
is a real result and must be written down, not quietly dropped.*
- **`ColumnMemoMiss` in the neighbourhood of 2030 % of all lookups, and staying there** ⇒
**confirmed**: the table is evicting on collision every Z plane. The fix is then obvious and cheap
(direct-indexed box keyed like `GSurfColCache`, or a set-associative table), and gets its own
task with the before/after this run supplies.
Report **both numbers**, not the ratio alone — a ratio cannot distinguish "few lookups" from "many".
## Notes for the reviewer (Claude)
- Confirm the miss counter sits inside the `if`, and the hit counter in an `else` — not computed
from a subtraction, which would silently agree with itself.
- Confirm `GetColumn`'s early-out path (if any is added later) cannot skip both counters.
- Confirm nothing else in `VoxelDensityOpStack.cpp` changed. `git diff --stat` should show one file
and a handful of lines.
+55
View File
@@ -3375,3 +3375,58 @@ covers both sites.
2. `e002bd4` (VerticalShafts connector capsules) — written, **not yet built**. Read
`Box verdicts over 60 VerticalShafts tiles`; 0 is the number to beat, `violations` must stay 0.
3. PERF on the op path; the warp squeeze stays parked.
## 2026-08-16 (b) — PERF: the suspects don't share an archetype, and the A/B needs no code
Orchestration pass on the parked PERF item. Three findings, none of them measured yet — they are
written here so the measurement that follows is aimed rather than exploratory.
### 1. The perf A/B requires NO new code — but it requires task 001 to be readable
`VoxelForge_ClassifyTile` and `VoxelForge_GenerateMesh` already exist at the call site. Since the
world is deterministic, the same seed and the same flight path generate the same tiles, so **two
Insights traces — `bUseOperatorStack` off, then on — are a clean A/B.** No instrument needed.
**But the totals cannot be compared directly.** With the stack on, tiles are *skipped*, so
`GenerateMesh` runs fewer times: a total-time comparison conflates "cheaper per tile" with "fewer
tiles", and those two point in opposite directions. **`TilesMeshed` from `CODEX-TASK-001` is the
denominator that separates them.**
⇒ **001 is a PREREQUISITE for the perf work, not a parallel item.** The handoff listed them as
independent (1 and 3); they are sequential. Order is now 001 → traces → attribution → fix.
### 2. "The op path is slower" is one number over two unrelated cost structures
Per archetype the suspects do not overlap, so a single figure cannot be acted on — the same lumping
error as the tile-skip counter, one level up:
- **SurfaceWorld** — the hashed column memo (below). Nothing to do with dispatch.
- **TunnelNetwork** — **19 virtual calls per voxel**: 16 from `BuildTunnelNetworkStack` (rock, room
graph, carve, the twelve detail mods, the worm) plus 3 from `AppendStructuralPost` (spine, seal,
passage carve). None inlinable, against one monolithic `GetDensityWithParams` the compiler can
inline and vectorise. Plus the known 12× gate re-test (stage B5's deliberate trade).
The measurement must therefore be taken **per archetype, one strate ticked at a time.**
### 3. The column memo hypothesis — `CODEX-TASK-002` written to test it
`FSurfaceColumnSource::GetColumn` uses a **direct-mapped 4096-entry hashed table**; `GSurfColCache`
uses a **direct-indexed box** with a `Computed[]` flag and therefore has no collisions at all.
The premise that makes the difference bite was checked rather than assumed: the mesher pre-samples
with **Z OUTERMOST** (`VoxelMarchingCubesMesher.cpp` ~226), sweeping a full XY plane per Z level, so
every column is revisited ~34 times per tile.
The table's own comment sized it against this and concluded 4096 entries "covers four chunks". **The
gap in that reasoning: a direct-mapped table evicts on collision, not on fullness.** ~1156 columns
in 4096 slots is a load factor of 0.28, at which ~285 columns (~25 %) still share a slot with
another. Those recompute the entire height stack — cliff's four structural resamples included — on
*every* Z plane. ~1156 column computations on the original path vs ~10 000 on the op path: **~9×**,
on the most expensive archetype in the plugin.
**This is a derivation, not a measurement.** Six confident chains in this project have reversed on
an unchecked premise, so `CODEX-TASK-002` adds `ColumnMemoHit` / `ColumnMemoMiss` and **fixes
nothing** — deliberately, so the instrument and the fix cannot land in the same build and make each
other unreadable. A negative result is specified as a real result.
`CODEX-TASK-002-column-memo-thrash.md`. **Nothing was built or run.**