Skip to content

ng1: add resident lifecycle and telemetry bounds (M7+M8) - #758

Merged
techomancer merged 5 commits into
techomancer:mainfrom
Tzeentchnet:feature/gpu-upstream-m7-m8
Nov 17, 2026
Merged

techomancer merged 5 commits into
techomancer:mainfrom
Tzeentchnet:feature/gpu-upstream-m7-m8

Conversation

@Tzeentchnet

@Tzeentchnet Tzeentchnet commented Nov 16, 2026 •

Copy link
Copy Markdown
Contributor

Important

#757 is integrated as 30fb9840. This branch is rebased onto upstream main at a4848ef7 and now contains only the five M7+M8 commits.

Summary

  • add a stopped, quiescent lifecycle boundary that excludes producers, joins display/GFIFO workers, and snapshots registers plus both authoritative planes atomically
  • restore into a new resident epoch only after both plane halves validate and upload; invalidate stale captures/bindings and fail closed on missing halves, upload failure, device loss, or incomplete HostRW state
  • carry the same checkpoint contract through machine snapshots, CI rollback, debugger checkpoints, reset, and a second Newport head
  • add checked aggregate telemetry for primitive/mode/coverage/HostRW distributions, batches/flush reasons, GFIFO occupancy/backpressure, observations, and frame latency
  • add a 4 ms maximum batch age without splitting an architectural operation or bounded-set wait; keep profiling opt-in and map only one 64-byte metadata record per nonempty batch

Efficiency evidence and limits

This responds to the concern that efficient wgpu acceleration of REX3 may not be possible by making the costs visible and bounded, not by claiming a speedup.

A fresh independent mixed-workload rerun on both backends accepted 30 operations with zero unsupported work and produced:

  • Vulkan: 6 nonempty batches; conservative frame-cut p50/p95 bounds <=0.131/1.049 ms; presentation p50/p95 <=2.097/4.194 ms
  • D3D12: 7 nonempty batches; frame-cut p50/p95 <=0.066/2.097 ms; presentation p50/p95 <=4.194/4.194 ms
  • both: zero GFIFO backpressure/full/busy events, one attributed 8 MiB HostR map, zero composed readback, and zero profiling maps in the default path

The committed PROM-replay measurements remain test-path evidence, not production application timing. The optional local recorder input is not in the repository, so it was not used as fresh PR evidence. The mixed gate also includes GPU setup and a CPU oracle, so no host-CPU or end-to-end speed claim is made; production IRIX application acceptance remains open.

Scope

This PR adds five M7+M8 commits directly above merged #757: four milestone commits plus the post-CI corpus reconciliation. Production activation and frontends remain M9/M10 work.

Validation

The exact 130-case resident gate passed on both required hardware backends on an NVIDIA RTX 5070:

  • 43 resident CPU tests
  • 4 flat-fill, 2 line, 2 HostW walker, 1 mixed HostW interpreter/JIT, 2 HostR port/retirement, 1 shade-DDA, 1 pattern-state, 3 command-ABI, and 3 direct-state tests
  • 24 resident GPU + 10 direct scanout tests on Vulkan
  • 24 resident GPU + 10 direct scanout tests on D3D12

Additional checks passed:

  • cargo check --locked -p iris --no-default-features
  • cargo check --locked -p iris --no-default-features --features wgpu
  • cargo check --locked -p iris --no-default-features --features rex-gpu
  • cargo check --locked -p iris-gui --features iris/wgpu
  • upstream dual-Newport placement/window test
  • shared MGRAS wgpu fill regression after the common batch changes
  • patch-equivalent range-diff after rebasing onto main at a4848ef7 following ng1: add resident HostR and explicit observers (M5+M6) #757's squash merge

@Tzeentchnet
Tzeentchnet force-pushed the feature/gpu-upstream-m7-m8 branch 2 times, most recently from f5e9d14 to 42a905b Compare November 17, 2026 02:09
@Tzeentchnet
Tzeentchnet force-pushed the feature/gpu-upstream-m7-m8 branch from 42a905b to c77976a Compare November 17, 2026 03:06
@Tzeentchnet
Tzeentchnet force-pushed the feature/gpu-upstream-m7-m8 branch from c77976a to 7523a9f Compare November 17, 2026 03:34
@techomancer
techomancer merged commit eb82d0a into techomancer:main Nov 17, 2026
9 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants