Unity / C# / D3D12

Shader Hitch Pipeline

Unity graphics-state tracing and warmup, with a measured repair for repeated driver checks and full-route evidence from official Boat Attack and URP scenes.

Public source · Draft PR 4; native Boat Attack and URP evidence published; not merged

ENGINEERING CASE STUDY · DRAFT PR

Finding the work behind a half-second stall

Boat Attack exposed an avoidable cost in my warmup receipt path: each repeated environment check rehashed about 127 MB of loaded driver files on the main thread. Reusing the byte-attested snapshot removed that repeated work while retaining identity checks.

~563 → 0.27–0.42 msRepeated driver checks in native diagnostics

16 / 16Repaired Boat Attack full-route processes completed

0 above 200 msSampled CPU Update intervals in that repaired cohort

The first full byte hash still costs about 574 ms. The diagnostic processes and formal comparison cohorts are separate; these figures do not imply a general scheduler or GPU speedup.

1. Observe the whole route

   The initial warmup arms showed 583–629 ms maximum Update intervals. First-pass P99 alone concealed this regression.

2. Attribute the overhead

   Receipt writing synchronously rehashed four loaded driver DLLs. Native warmup batch timings did not include that extra CPU work.

3. Retain checks and reuse work

   A process-local snapshot preserves byte hashes. Each capture checks device, module and file metadata; changed or failed probes invalidate the snapshot.

4. Repeat the original workload

   Four processes per policy completed the original Boat Attack routes. Both the initial negative cohort and the repaired cohort remain inspectable.

Boat Attack before and after the repair

Largest sampled CPU Update interval across four processes per policy

Policy	Before repair (ms)	After repair (ms)
Disabled	47.089	58.353
All at once	591.826	51.737
Scheduled	629.288	57.728
Observed budget	594.053	90.256

These are two sequential cohorts with different frozen Players and plans. Caches were retained. The before/after comparison supports a diagnosed overhead repair; the cohorts are not pooled as one unchanged experimental cell.

Inspect all 32 Boat Attack runs

All 32 Boat Attack process points; rounded to three decimals

Cohort	Run	Policy	First-pass P99 (ms)	Whole-run max (ms)
Before	1	Disabled	35.448	47.089
Before	2	All at once	34.032	584.435
Before	3	Observed budget	34.735	594.053
Before	4	Scheduled	32.926	629.288
Before	5	All at once	32.285	583.276
Before	6	Scheduled	36.859	584.297
Before	7	Disabled	34.352	44.595
Before	8	Observed budget	36.099	587.168
Before	9	Scheduled	34.716	595.050
Before	10	Observed budget	34.470	593.155
Before	11	All at once	36.729	585.566
Before	12	Disabled	35.140	43.982
Before	13	Observed budget	33.526	593.747
Before	14	Disabled	36.984	45.609
Before	15	Scheduled	35.673	584.487
Before	16	All at once	36.785	591.826
After	1	Disabled	40.516	58.353
After	2	All at once	37.707	49.883
After	3	Observed budget	37.879	90.256
After	4	Scheduled	39.266	50.172
After	5	All at once	38.459	46.435
After	6	Scheduled	39.902	57.728
After	7	Disabled	41.849	57.134
After	8	Observed budget	39.530	59.639
After	9	Scheduled	35.917	49.104
After	10	Observed budget	37.744	45.781
After	11	All at once	36.532	50.225
After	12	Disabled	38.869	46.289
After	13	Observed budget	38.189	49.239
After	14	Disabled	36.329	46.063
After	15	Scheduled	38.987	56.373
After	16	All at once	38.269	51.737

Download original points with full precision and provenance hashes

Four official URP scenes, with the outlier retained

Terminal, Garden, Oasis and Cockpit completed their original timelines in all 16 formal processes. Each policy used the same final Player. Outcomes remain mixed: the observed-budget arm includes an unexplained 381.3446 ms sampled Update interval.

Inspect all 16 URP runs

All 16 URP process points; all original scene timelines retained

Run	Policy	Observer time (s)	Maximum Update (ms)
1	Disabled	264.745125	212.1705
2	All at once	264.896429	218.3915
3	Observed budget	264.854141	217.9709
4	Scheduled	265.136408	214.2184
5	All at once	264.579465	217.7115
6	Scheduled	264.819980	216.0261
7	Disabled	264.827879	220.8122
8	Observed budget	265.147337	381.3446
9	Scheduled	264.695923	217.3337
10	Observed budget	264.916918	218.6572
11	All at once	265.009663	221.4843
12	Disabled	264.573533	220.9166
13	Observed budget	264.792184	219.5831
14	Disabled	264.866628	217.6925
15	Scheduled	264.678605	212.8079
16	All at once	264.669977	217.3270

Download full-precision points, per-scene values and hardware conditions

What this result establishes

A specific repair to instrumentation overhead, and completed native integration on the recorded Boat Attack and URP routes. Overall scheduling speedup, useful first-draw coverage gains and Megacity acceptance remain unproven.

Windows · Unity 6000.1.0f1 · D3D12 / IL2CPP · AMD Radeon AI PRO R9700 · application, OS and driver caches retained. Figures and points are historical records, separate from this website’s live WebGPU lab.

All published diagnostic timings · Figure and data sources · Choose a role-specific CV

Work and recorded results

• Reduced repeated driver-attestation checks from ~563 ms to 0.27-0.42 ms in native Unity/D3D12 diagnostics; the repaired Boat Attack cohort completed 16/16 full routes. Draft PR; overall scheduler speedup remains unproven.

Scope of the evidence

• Draft PR 4 is not merged. Its published reports retain their own frozen Player, adapter and protocol identities; the current source link is not a claim that all native runs used this exact source revision.
• The initial full driver hash still costs about 574 ms. Reuse checks device, module and file metadata and invalidates changed identities.
• Boat Attack retains both 16-process cohorts (32 total); the repaired cohort recorded no sampled CPU Update interval above 200 ms. The separate URP cohort completed 16 processes across four original scene timelines. These are one Windows D3D12/IL2CPP hardware cell with caches retained, not GPU frame-time or general scheduler-speedup evidence.
• The URP 381.3446 ms interval remains unexplained. Useful first-draw coverage and overall scheduling speedup remain unproven; Megacity Main, six SubScenes and final Player acceptance are incomplete.
• The retained historical 384-object showcase reported cold-reveal P95 35.706 to 4.210 ms and about 20% longer total scheduled warmup; its image and numbers describe a separate historical workload.

Source and evidence

Reviewed source revision: 9c7fc7b75fa7ac1bca2e407821414cf5941e9b8c

Public source at this revision ↗

View draft PR (not merged) ↗

Current main: 6218483755b40f6fd3742754b9e71c53b74d3bd4. The draft evidence above is separate from this merged version.

• Repeated-check repair and Boat Attack cohorts
• Official URP four-scene full-timeline evidence
• Incomplete Megacity acceptance
• Separate historical showcase conditions

Updated 2026-09-14. CV wording and this page use the same project manifest.