research(vm0): Campaign AR's pack-off path IS the pre-campaign renderer - PASS

Exact-pixel and production-perf comparison of 6c79d35c (+ only the A2
normal files, so no terrain mask is needed) against c51b07ef with no pack:
connected as +Acdream, visible window, one isolated config clone per
variant, pinned clocks. Open field: the only differences are idle pose,
mana digits and a passing flyer. Holtburg: same-binary-twice defines the
dynamic mask (9.8%); in the static 90% both self-diffs have ZERO pixels
with |d|>=8 while base+normals vs HEAD-off has 841/729 - all streaks inside
the animated lifestone. Buildings, ground, trees, sky and UI are clean.

Perf (uncapped Release, no automation observer, ACDREAM_FRAME_PROF=1):
Holtburg CPU p50 4.7 -> 4.1 ms, Arwic 6.0 -> 5.2 ms, GPU unchanged,
alloc/frame 574 KB -> 21 KB. No regression; F5b's '27.8 ms retail CPU' was
the observer.

Three false alarms recorded so nobody repeats them: the isolated gate
settings lack fieldOfView (90 vs the real 86.33 -> a 0.952 zoom); the real
%APPDATA% settings still selected acdream.atmospheric/low (pack ON); a
minimized GLFW window is throttled and never settles.

Tools: tools/vm0/capture-visible.ps1 (pre-campaign gate + -Exe/-Live/
-ConfigDir/-CharacterName/-PreCaptureCommand), tools/vm0/perf-run.sh,
and -BuildingDetailTextures on run-offline-pixel-gate.ps1. Baseline
patches under docs/research/evidence/vm0/.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
Erik 2026-08-22 21:40:40 +02:00
parent 4a69ecd32e
commit 30e72a2af8
12 changed files with 721 additions and 2 deletions

View file

@ -76,7 +76,7 @@ single-pass path darkens by ~10 %. Neither explains a scene-wide level.
| Slice | Closes | Gate |
|---|---|---|
| VM0 | F1 — default-path invariance | automated + Fable report |
| VM0 | F1 — default-path invariance | **CLOSED 2026-08-22 — PASS**; [report](../research/2026-08-22-vm0-default-path-invariance.md) |
| VM1 | F2 — invented detail fade | automated |
| VM2 | F3 — which retail detail path ran | **CLOSED 2026-08-22** — single-pass; see [cdb note](../research/2026-08-22-vm2-retail-detail-path-cdb.md) |
| VM3 | F4 — linear-light post stack | automated + **owner visual** |
@ -85,7 +85,7 @@ single-pass path darkens by ~10 %. Neither explains a scene-wide level.
| VM6 | Foliage wind | automated + **owner visual** |
| VM7 | Closeout: full gates, register, roadmap, merge | **owner** |
Order is VM2 (done) → VM0 → VM1 → VM4 → VM5 → VM3 → VM6 → VM7. VM0 goes first because everything after it is measured
Order is VM2 (done) → VM0 (done) → VM1 → VM4 → VM5 → VM3 → VM6 → VM7. VM0 goes first because everything after it is measured
against the baseline it establishes.
---
@ -126,6 +126,8 @@ performance against `6c79d35c`.
6. Add the masked comparison as a repeatable tool mode
(`-Baseline … -TerrainMask`) so VM3/VM6 can re-run it.
**OUTCOME (2026-08-22): PASS.** Connected, visible-window, isolated-config captures: base+normals vs HEAD-off has zero strong static differences outside the animated lifestone at Holtburg and the open field; production CPU is 13% cheaper and allocation 25x lower on HEAD's pack-off path (no observer, uncapped Release). Three config traps (isolated FOV, the real Roaming settings still selecting the pack, minimized-window throttling) are recorded in the report. Replaced the planned terrain mask with a base+normals baseline (no mask needed).
**Acceptance:** the two pixel assertions in step 4 hold; the perf A/B shows
no default-path regression; the Fable report records the exact commands,
commits and artifact paths. If a regression is found, it is fixed here and

View file

@ -0,0 +1,119 @@
# VM0 — is Campaign AR's pack-off path the pre-campaign renderer?
**Date:** 2026-08-22 · **Campaign:** VM slice VM0 · **Status:** CLOSED — PASS
**Question:** with no render pack selected, does `c51b07ef` (Campaign AR) draw
the same pixels and cost the same as `6c79d35c` (pre-campaign main), outside
the two intended parity changes (A2 terrain normals, #226 building detail)?
**Answer:** yes, to the measured noise floor — and the pack-off path is
faster and allocates less.
## Binaries
| Name | Source | Notes |
|---|---|---|
| `base` | `6c79d35c` | + `evidence/vm0/baseline-configdir.patch` (HEAD's 13-line `ACDREAM_CONFIG_DIR/DATA_DIR/CACHE_DIR` hunk in `ApplicationPathSet`, path resolution only) so it can run from an isolated config |
| `base+normals` | `6c79d35c` + `evidence/vm0/baseline-normals.patch` (only `LandblockMesh.cs` + `TerrainVertex.cs` from the campaign) + the config-dir hunk | isolates A2 so the key comparison needs no terrain mask |
| `HEAD` | `c51b07ef` | `BuildingDetailTextures` off and on |
All Release. Worktrees `.claude/worktrees/vm0-base` and `vm0-base-normals`
were throwaway; the two patches reproduce them.
## Method that finally worked (and the three that did not)
Every run: connected to the local ACE as `+Acdream` (name-selected via
`--session-config`; ACE reorders the roster by last login, so an index is not
stable), **visible normal window** (the product condition), one **isolated
config clone** per variant under `artifacts/vm0/cfg-*` with
`renderPack = retail/off`, pinned `ACDREAM_DAY_GROUP=0`,
`ACDREAM_WORLD_TIME=0.5`, `ACDREAM_SKY_PHASE_SECONDS=0`, `ACDREAM_MSAA_SAMPLES=0`,
1280×720, 12 s settle, one screenshot. Helper:
`tools/vm0/capture-visible.ps1` (the `6c79d35c` pixel-gate script with `-Exe`,
`-Live`, `-ConfigDir`, `-CharacterName`, `-PreCaptureCommand`).
Three earlier attempts produced 8891 % pixel differences that were **all
configuration, none renderer** — recorded so nobody repeats them:
1. **Hidden-window captures vs old-tool captures differed by a 0.952 zoom.**
HEAD's gate writes an isolated `settings.json` with no `fieldOfView`, so
HEAD rendered at the default 90° gameFOV while the old binary read the real
file's `86.33°`. Not a renderer change. Consequence: captures from the new
`run-offline-pixel-gate.ps1` and from the pre-campaign tool are **not
comparable** to each other; each tool is self-consistent.
2. **Connected HEAD frames were brighter with tree shadows.** The real Roaming
`settings.json` still held `renderPack = acdream.atmospheric/low` from the
owner's live gate — the pack was ON. (The reviewer had been reading and
editing a stale `%LOCALAPPDATA%\acdream\settings.json`; the client's
Windows config root is `%APPDATA%\acdream`.) Every HEAD run must pin the
selection explicitly.
3. **A minimized HEAD window rendered brighter than hidden/visible.** Windows
throttles an iconified GLFW surface; the 12 s settle had not completed.
Minimized is not a product condition and is not used.
## Pixel results
Exact per-pixel comparison (`max |ΔRGB|`), no tolerance.
**Open field (logout spot, 60.7S 89.1W, few dynamics):**
| Pair | px differing | notes |
|---|---|---|
| base vs base+normals (A2 only) | 12.7 %, mean Δ 1.5 | terrain shading, low amplitude |
| **base+normals vs HEAD-off (KEY)** | **4.3 %, mean Δ 9** | all of it: idle-pose outline, regenerating mana digits, one ambient flyer crossing the camera — ground, trees and UI clean (`diff-I: HEAD-off vs base+normals (KEY).png`) |
| HEAD-off vs HEAD-on (detail) | 3.3 % | no buildings in view; same noise class |
**Holtburg (`/telepoi Holtburg`, buildings, NPCs, animated lifestone):**
the same binary twice differs in 2535 % of pixels by ±17 (sun/time
sub-steps), so a raw count is meaningless. Pixels with |Δ| ≥ 8 between two
logins of the *same* binary define the dynamic mask (lifestone, NPCs, particle
emitter, radar blips, digits; 9.8 % of the frame after a 6 px dilation). In
the remaining 90.2 %:
| Pair | |Δ| ≥ 8 px in static region | max |
|---|---|---|
| base+normals self (two logins) | **0** | 7 |
| HEAD-off self (two logins) | **0** | 7 |
| **base+normals vs HEAD-off (KEY, run 1)** | **841 (0.10 %)** | 128 |
| **base+normals vs HEAD-off (KEY, run 2)** | **729 (0.09 %)** | 128 |
| base vs base+normals (A2 only) | 7,884 (0.95 %) | 212 |
| HEAD-off vs HEAD-on (detail) | 513 (0.06 %) | 255 |
The 841/729 KEY pixels are streaks inside the animated lifestone crystal that
escaped the mask dilation (`evidence/vm0/holtburg-KEY-static-strong-diff.png`);
buildings, roofs, ground, trees, sky and every UI element are clean. The A2
row is the expected shading change on terrain. The detail row is below the
threshold on building shells because the fallback blend's mean factor is
1.033 (+3 %); VM1's single-pass re-port (10 %) will make it measurable.
Evidence: `docs/research/evidence/vm0/*.png`, full captures under
`artifacts/vm0/` (not committed).
## Production performance (no observer)
`tools/vm0/perf-run.sh`: connected, **uncapped Release**, no
`ACDREAM_AUTOMATION_ARTIFACT_DIR` (so no render-scene observer),
`ACDREAM_FRAME_PROF=1`, same isolated config, 75 s at each spot; last six
5-second `[frame-prof]` windows (`evidence/vm0/perf-frame-prof.txt`):
| Spot | Binary | CPU p50 / p95 / p99 | GPU p50 / p95 | alloc KB/frame | Gen0 per 5 s |
|---|---|---|---|---|---|
| Holtburg | base | 4.7 / 5.1 / 5.5 ms | 0.4 / 0.5 | 574 | 1213 |
| Holtburg | **HEAD off** | **4.1 / 4.4 / 4.8 ms** | 0.4 / 0.5 | **21** | 01 |
| Arwic (dense) | base | 6.0 / 6.5 / 7.0 ms | 0.7 / 0.8 | 582 | 10 |
| Arwic (dense) | **HEAD off** | **5.2 / 5.6 / 6.1 ms** | 0.7 / 0.9 | **29** | 01 |
No regression. The pack-off path is ~13 % cheaper on CPU with ~25× less
managed allocation per frame; GPU unchanged. This also settles review F5b:
the "27.8 ms retail-path CPU" in the Stage-2 connected report was the
automation observer, not the product.
## Verdict
VM0 PASSES. Campaign AR's default path is the pre-campaign renderer plus the
two declared parity changes. The F1 blocker is cleared.
## Carried forward
- The new pixel gate's isolated settings should carry `fieldOfView` (and
`gamma`) so its captures mean the same as the product's — filed for VM7's
tool tidy; not a renderer issue.
- `run-offline-pixel-gate.ps1` gained `-BuildingDetailTextures` (VM0).

View file

@ -0,0 +1,31 @@
diff --git a/src/AcDream.Platform/ApplicationPathSet.cs b/src/AcDream.Platform/ApplicationPathSet.cs
index a6743ad1..76e3a1f5 100644
--- a/src/AcDream.Platform/ApplicationPathSet.cs
+++ b/src/AcDream.Platform/ApplicationPathSet.cs
@@ -72,6 +72,16 @@ public sealed record ApplicationPathSet(
{
platform ??= ApplicationPathEnvironment.Instance;
+ // Explicit method/CLI arguments remain authoritative. The environment
+ // seam lets graphical automation and portable installations isolate
+ // all mutable user state without rewriting the real user's settings.
+ configDirectory ??= NonEmpty(
+ platform.GetEnvironmentVariable("ACDREAM_CONFIG_DIR"));
+ dataDirectory ??= NonEmpty(
+ platform.GetEnvironmentVariable("ACDREAM_DATA_DIR"));
+ cacheDirectory ??= NonEmpty(
+ platform.GetEnvironmentVariable("ACDREAM_CACHE_DIR"));
+
string config;
string data;
string cache;
@@ -140,6 +150,9 @@ public sealed record ApplicationPathSet(
return Path.Combine(root, leaf);
}
+ private static string? NonEmpty(string? value) =>
+ string.IsNullOrWhiteSpace(value) ? null : value;
+
private static string RequireFolder(
IApplicationPathEnvironment platform,
Environment.SpecialFolder folder)

View file

@ -0,0 +1,149 @@
diff --git a/src/AcDream.Core/Terrain/LandblockMesh.cs b/src/AcDream.Core/Terrain/LandblockMesh.cs
index 81e67249..e3a09fca 100644
--- a/src/AcDream.Core/Terrain/LandblockMesh.cs
+++ b/src/AcDream.Core/Terrain/LandblockMesh.cs
@@ -56,27 +56,21 @@ public static class LandblockMesh
throw new ArgumentException("heightTable must have 256 entries", nameof(heightTable));
// Pre-sample all 81 heights into a 2D array (x-major indexing). This
- // doubles as the source for per-vertex normals via central differences
- // (Phase 3b lighting, preserved through the per-cell refactor).
+ // is also the source for retail's topology-aware vertex normals.
var heights = new float[HeightmapSide, HeightmapSide];
for (int x = 0; x < HeightmapSide; x++)
for (int y = 0; y < HeightmapSide; y++)
heights[x, y] = heightTable[block.Height[x * HeightmapSide + y]];
- // Pre-compute all 81 vertex normals so the inner cell loop is a pure
- // lookup. Central differences on the heightmap → smooth normal field.
- var normals = new Vector3[HeightmapSide, HeightmapSide];
- for (int x = 0; x < HeightmapSide; x++)
- for (int y = 0; y < HeightmapSide; y++)
- {
- int xL = Math.Max(x - 1, 0);
- int xR = Math.Min(x + 1, HeightmapSide - 1);
- int yD = Math.Max(y - 1, 0);
- int yU = Math.Min(y + 1, HeightmapSide - 1);
- float dx = (heights[xR, y] - heights[xL, y]) / ((xR - xL) * CellSize);
- float dy = (heights[x, yU] - heights[x, yD]) / ((yU - yD) * CellSize);
- normals[x, y] = Vector3.Normalize(new Vector3(-dx, -dy, 1f));
- }
+ // Retail CLandBlockStruct::calc_lighting accumulates the normalized
+ // plane normal of every incident terrain polygon at each of the 81
+ // shared height-sample vertices, then normalizes the sum. Use the same
+ // split hash and triangle topology as the emitted mesh; this changes
+ // lighting only, never positions, indices, or the collision surface.
+ var normals = BuildRetailVertexNormals(
+ heights,
+ landblockX,
+ landblockY);
var vertices = new TerrainVertex[VerticesPerLandblock];
var indices = new uint[VerticesPerLandblock]; // 1 index per vertex (no deduplication)
@@ -173,6 +167,85 @@ public static class LandblockMesh
return new LandblockMeshData(vertices, indices);
}
+ private static Vector3[,] BuildRetailVertexNormals(
+ float[,] heights,
+ uint landblockX,
+ uint landblockY)
+ {
+ var normalSums = new Vector3[HeightmapSide, HeightmapSide];
+
+ for (int cy = 0; cy < CellsPerSide; cy++)
+ {
+ for (int cx = 0; cx < CellsPerSide; cx++)
+ {
+ var posBL = new Vector3( cx * CellSize, cy * CellSize, heights[cx, cy ]);
+ var posBR = new Vector3((cx + 1) * CellSize, cy * CellSize, heights[cx + 1, cy ]);
+ var posTR = new Vector3((cx + 1) * CellSize, (cy + 1) * CellSize, heights[cx + 1, cy + 1]);
+ var posTL = new Vector3( cx * CellSize, (cy + 1) * CellSize, heights[cx, cy + 1]);
+
+ var split = TerrainBlending.CalculateSplitDirection(
+ landblockX, (uint)cx, landblockY, (uint)cy);
+
+ if (split == CellSplitDirection.SWtoNE)
+ {
+ AccumulateFaceNormal(
+ normalSums,
+ posBL, cx, cy,
+ posBR, cx + 1, cy,
+ posTR, cx + 1, cy + 1);
+ AccumulateFaceNormal(
+ normalSums,
+ posBL, cx, cy,
+ posTR, cx + 1, cy + 1,
+ posTL, cx, cy + 1);
+ }
+ else
+ {
+ AccumulateFaceNormal(
+ normalSums,
+ posBL, cx, cy,
+ posBR, cx + 1, cy,
+ posTL, cx, cy + 1);
+ AccumulateFaceNormal(
+ normalSums,
+ posBR, cx + 1, cy,
+ posTR, cx + 1, cy + 1,
+ posTL, cx, cy + 1);
+ }
+ }
+ }
+
+ var normals = new Vector3[HeightmapSide, HeightmapSide];
+ for (int x = 0; x < HeightmapSide; x++)
+ {
+ for (int y = 0; y < HeightmapSide; y++)
+ {
+ Vector3 sum = normalSums[x, y];
+ normals[x, y] = sum.LengthSquared() > 0f
+ ? Vector3.Normalize(sum)
+ : Vector3.UnitZ;
+ }
+ }
+
+ return normals;
+ }
+
+ private static void AccumulateFaceNormal(
+ Vector3[,] normalSums,
+ Vector3 p0, int x0, int y0,
+ Vector3 p1, int x1, int y1,
+ Vector3 p2, int x2, int y2)
+ {
+ Vector3 cross = Vector3.Cross(p1 - p0, p2 - p0);
+ if (cross.LengthSquared() <= 0f)
+ return;
+
+ Vector3 faceNormal = Vector3.Normalize(cross);
+ normalSums[x0, y0] += faceNormal;
+ normalSums[x1, y1] += faceNormal;
+ normalSums[x2, y2] += faceNormal;
+ }
+
private static void WriteCell(
TerrainVertex[] verts, ref int vi,
uint d0, uint d1, uint d2, uint d3,
diff --git a/src/AcDream.Core/Terrain/TerrainVertex.cs b/src/AcDream.Core/Terrain/TerrainVertex.cs
index a031e837..69775c23 100644
--- a/src/AcDream.Core/Terrain/TerrainVertex.cs
+++ b/src/AcDream.Core/Terrain/TerrainVertex.cs
@@ -11,11 +11,11 @@ namespace AcDream.Core.Terrain;
/// which of the 4 cell corners a given vertex represents from
/// <c>gl_VertexID % 6</c> plus the split direction bit.
///
-/// Normal is stored per vertex via Phase 3b's central-difference scheme on
-/// the 9×9 heightmap — this lets the fragment shader interpolate a smooth
-/// normal across triangles (softer than WorldBuilder's <c>dFdx</c>/<c>dFdy</c>
-/// flat-shaded approach). UVs are derived from the corner index in the
-/// vertex shader — not stored here.
+/// Normal is stored per vertex using retail's terrain-lighting rule: each
+/// shared height-sample vertex receives the normalized plane normals of its
+/// incident, split-aware terrain triangles and normalizes their sum. The
+/// fragment shader interpolates that smooth result across triangles. UVs are
+/// derived from the corner index in the vertex shader — not stored here.
///
/// Size: 12 (position) + 12 (normal) + 4*4 (Data0..3) = 40 bytes.
/// </summary>

View file

@ -0,0 +1,31 @@
import sys, numpy as np
from PIL import Image
root = r"artifacts/vm0"
def load(n): return np.asarray(Image.open(f"{root}/{n}/screenshots/world-offline.png").convert("RGB")).astype(np.int16)
def cmp(a, b, name, maskTop=0):
A, B = load(a), load(b)
assert A.shape == B.shape, (A.shape, B.shape)
d = np.abs(A - B).max(axis=2)
H, W = d.shape
region = d[maskTop:, :]
diff = region > 0
n = int(diff.sum()); tot = region.size
rows = np.where(diff.any(axis=1))[0]
print(f"{name}: {n} / {tot} px differ ({100.0*n/tot:.5f}%)"
+ (f" rows {rows.min()+maskTop}-{rows.max()+maskTop}, max|d|={int(region.max())}, mean|d| over differing={region[diff].mean():.2f}" if n else " IDENTICAL"))
if n:
hist = np.bincount(region[diff].astype(np.int64), minlength=8)
print(f" |d| histogram 1..7+: {hist[1:8].tolist()} ...>=8: {int((region[diff]>=8).sum())}")
vis = np.zeros((H, W, 3), np.uint8)
vis[..., 0] = np.clip(d * 16, 0, 255).astype(np.uint8) # red = difference amplified x16
base = (A.astype(np.float32) * 0.25).astype(np.uint8)
vis = np.maximum(vis, base)
Image.fromarray(vis).save(f"{root}/diff-{name}.png")
return n
print("== whole frame, exact ==")
cmp("base", "base-normals", "A2-normals-only")
cmp("base-normals", "head-off", "HEAD-off_vs_base+normals")
cmp("head-off", "head-on", "detail-on_vs_off")
cmp("base", "head-off", "HEAD-off_vs_base")
print("== top-280 sky band masked (the tool's documented self-noise band) ==")
cmp("base-normals", "head-off", "HEAD-off_vs_base+normals_masked", 280)

Binary file not shown.

After

Width:  |  Height:  |  Size: 418 KiB

Binary file not shown.

After

Width:  |  Height:  |  Size: 545 KiB

Binary file not shown.

After

Width:  |  Height:  |  Size: 876 KiB

View file

@ -0,0 +1,4 @@
holtburg-base: p50=4.7 p95=5.1 p99=5.8 max=8.8 | gpu_ms p50=0.4 p95=0.5 | alloc_kb p50=574.4 max=600.3 gc=12/3/0;p50=4.7 p95=5.0 p99=5.3 max=6.1 | gpu_ms p50=0.4 p95=0.5 | alloc_kb p50=574.4 max=689.8 gc=13/4/0;p50=4.7 p95=5.0 p99=5.5 max=7.7 | gpu_ms p50=0.4 p95=0.5 | alloc_kb p50=574.4 max=689.8 gc=12/4/0;p50=4.7 p95=5.0 p99=5.3 max=6.2 | gpu_ms p50=0.4 p95=0.5 | alloc_kb p50=574.4 max=689.8 gc=12/3/0;p50=4.7 p95=5.1 p99=5.4 max=6.4 | gpu_ms p50=0.4 p95=0.5 | alloc_kb p50=574.4 max=623.1 gc=13/4/0;p50=4.7 p95=5.1 p99=5.8 max=10.1 | gpu_ms p50=0.4 p95=0.5 | alloc_kb p50=574.4 max=692.6 gc=12/3/0;
holtburg-head: p50=4.1 p95=4.4 p99=4.7 max=6.5 | gpu_ms p50=0.4 p95=0.5 | alloc_kb p50=20.9 max=47.2 gc=1/0/0;p50=4.1 p95=4.4 p99=4.7 max=5.5 | gpu_ms p50=0.4 p95=0.5 | alloc_kb p50=20.9 max=139.2 gc=1/0/0;p50=4.2 p95=4.4 p99=4.6 max=5.7 | gpu_ms p50=0.4 p95=0.5 | alloc_kb p50=20.9 max=156.5 gc=0/0/0;p50=4.1 p95=4.4 p99=4.7 max=5.3 | gpu_ms p50=0.4 p95=0.5 | alloc_kb p50=20.9 max=220.9 gc=1/0/0;p50=4.1 p95=4.4 p99=5.5 max=9.0 | gpu_ms p50=0.4 p95=0.5 | alloc_kb p50=20.9 max=67.0 gc=0/0/0;p50=4.2 p95=4.4 p99=4.6 max=5.4 | gpu_ms p50=0.4 p95=0.5 | alloc_kb p50=20.9 max=136.3 gc=1/0/0;
arwic-base: p50=6.0 p95=6.5 p99=7.0 max=8.9 | gpu_ms p50=0.7 p95=0.8 | alloc_kb p50=582.5 max=646.5 gc=10/3/0;p50=6.0 p95=6.4 p99=6.8 max=7.5 | gpu_ms p50=0.7 p95=0.8 | alloc_kb p50=582.5 max=697.7 gc=9/3/0;p50=6.0 p95=6.5 p99=7.0 max=7.6 | gpu_ms p50=0.7 p95=0.8 | alloc_kb p50=582.5 max=700.7 gc=10/3/0;p50=5.9 p95=6.4 p99=7.0 max=11.5 | gpu_ms p50=0.7 p95=0.8 | alloc_kb p50=582.5 max=700.9 gc=10/4/0;p50=6.0 p95=6.4 p99=7.0 max=11.7 | gpu_ms p50=0.7 p95=0.8 | alloc_kb p50=582.5 max=620.8 gc=10/3/0;p50=6.0 p95=6.5 p99=7.1 max=7.9 | gpu_ms p50=0.7 p95=0.8 | alloc_kb p50=582.5 max=732.5 gc=10/3/0;
arwic-head: p50=5.1 p95=5.5 p99=5.9 max=6.5 | gpu_ms p50=0.7 p95=0.9 | alloc_kb p50=28.9 max=68.1 gc=1/0/0;p50=5.2 p95=5.6 p99=6.3 max=8.4 | gpu_ms p50=0.7 p95=0.9 | alloc_kb p50=28.8 max=143.9 gc=1/0/0;p50=5.1 p95=5.5 p99=6.1 max=9.6 | gpu_ms p50=0.7 p95=0.9 | alloc_kb p50=28.9 max=143.9 gc=0/0/0;p50=5.1 p95=5.5 p99=5.9 max=6.3 | gpu_ms p50=0.7 p95=0.9 | alloc_kb p50=28.9 max=144.1 gc=1/0/0;p50=5.2 p95=5.6 p99=6.6 max=12.4 | gpu_ms p50=0.7 p95=0.9 | alloc_kb p50=28.8 max=68.2 gc=1/0/0;p50=5.2 p95=5.5 p99=5.9 max=6.6 | gpu_ms p50=0.7 p95=0.9 | alloc_kb p50=28.8 max=144.1 gc=0/0/0;