Skip to content

graphics: complete GPU milestones M9-M16 - #767

Merged
techomancer merged 1 commit into
techomancer:mainfrom
Tzeentchnet:feature/gpu-upstream-m9-m16
Nov 17, 2026
Merged

techomancer merged 1 commit into
techomancer:mainfrom
Tzeentchnet:feature/gpu-upstream-m9-m16

Conversation

@Tzeentchnet

@Tzeentchnet Tzeentchnet commented Nov 17, 2026 •

Copy link
Copy Markdown
Contributor

Important

#757 and #758 are integrated. This PR is rebased directly onto upstream main at eb82d0ae and contains exactly one M9-M16 commit.

Summary

  • activate Newport's resident wgpu path in production builds and wire explicit raster/presentation configuration through machines and frontends
  • complete IMPACT correctness, board-matrix, GPU-residency, telemetry, failure, and snapshot lifecycle work for Indigo2 and Octane
  • close VPro's silent framebuffer-fallback and transfer gaps, add typed telemetry, and add atomic stopped snapshots with GPU-authoritative restore/reseed on Octane
  • make GR2 XZ/Extreme complete schema-10 snapshot, rollback, reset, and stopped-lifecycle participants
  • add representative GR2 span/batching telemetry and retain gr2-jit after the predefined resident-GPU gate produced an evidence-based no-go

Milestones

M9+M10 - Newport production activation and frontends

  • compile and manage the resident worker through the production board lifecycle
  • keep direct wgpu and explicit readback compatibility fixed for the board lifetime; failures remain explicit and fail closed
  • validate and share the selected device across dual Newport heads and direct presentation
  • add configuration, machine, CLI, GUI, headless, screenshot, VNC/capture, and compatibility-mode wiring without silent mode changes

M11+M12 - IMPACT correctness, residency, and lifecycle

  • extend window/buffer, pixel-format, destination-RE, texture, context, fog, lighting, and Solid/High/Maximum coverage across the interpreter and gr4-jit
  • keep GPU pixel pages authoritative, attribute observations and transfer bytes, and turn GPU synchronization failures into sticky fail-closed state
  • snapshot complete stopped HQ3/HQ4, GE, RSS, ERAM, TRAM, board-memory, interrupt, and display state
  • integrate atomic IMPACT disk snapshots and rollback with Octane's XIO wrapper, including GPU mirror reseeding after restore

M13+M14 - VPro residency and Octane snapshots

  • reject unsupported production GPU work instead of silently drawing through stale CPU memory
  • move supported copy/view/depth operations into ordered GPU work and keep DMA/debug/screenshot observations bounded and reason-tagged
  • add batch, copy, readback, rejection, latency, and device-loss telemetry
  • snapshot command/display FIFOs, contexts, query/fence state, lookup tables, board state, and 32/128 MiB memory before publishing a restored epoch
  • support exact Octane V6/V8/V10/V12 round trips with atomic validation and GPU-authoritative memory reseeding

M15+M16 - GR2 lifecycle and measured GPU decision

  • add exact named state, vram, and zbuf parts while retaining schema 6-9 decoding
  • preserve HQ2 parser/GL continuation, RE3 state, SRAM/microcode, display state, FIN2 waits, and unread GEDMA across disk, rollback, and live checkpoints
  • measure checked-in GL traces: 480-1,116 spans, 160,740-177,600 pixels, maximum batchable runs of 300-816, and 3-11 synchronization boundaries
  • measure warmed gr2-jit at 178-568 Mpx/s across 4-1,280-pixel spans, 1.33-2.86x faster than the interpreter

Decision: defer a resident RE3 wgpu implementation and retain gr2-jit. The traces establish batching potential, but there is no GPU prototype that beats the warm-JIT large-span baseline or proves the <=5% short-span bound; ordinary presentation still scans 1,310,720 host pixels per frame, and no repeatable representative guest benefit has been measured. This does not conclude that efficient acceleration is impossible.

Commit scope

The incremental M9-M16 change is one commit (786a7306), touching 95 files with 7,016 insertions and 956 deletions. GitHub now displays only this isolated commit.

Validation

Current rebased tip on main at eb82d0ae:

  • exact 131-case Newport gate passed without skips on Vulkan and D3D12
    • Vulkan: NVIDIA GeForce RTX 5070
    • D3D12: NVIDIA GeForce RTX 5070
  • GR2 live/rollback and disk snapshot integration passed
  • Octane IMPACT and VPro disk snapshot round trips passed
  • D3D12 VPro flat-triangle depth parity and GPU restore/reseed passed with the backend explicitly selected
  • cargo check for the combined wgpu,gpujit core configuration passed
  • iris-gui with iris/wgpu passed
  • git diff --check passed

Full consolidated-campaign validation:

  • GR2: 125 active release tests passed; four ignored
  • snapshot/schema: 39 active tests passed; one ignored
  • IMPACT CPU/gr4-jit: 227 active tests passed; two ignored
  • shared IMPACT/VPro wgpu suite: 34 active tests passed and one ignored on each of Vulkan and D3D12
  • IMPACT and VPro GPU snapshot/reseed tests passed on both backends
  • GR2 six-sample interleaved span benchmark passed
  • all eight core/frontend feature checks passed

One existing-checkout caveat: the tracked VPro golden fixture was checked out as CRLF despite eol=lf, so the default byte-string comparison reported line-ending differences locally. The tolerance verifier reported zero moved rows and zero numeric differences in every field. A normalized checkout uses the tracked LF form; no broad fixture rewrite is included.

@Tzeentchnet
Tzeentchnet force-pushed the feature/gpu-upstream-m9-m16 branch 2 times, most recently from f5d26ce to 226a785 Compare November 17, 2026 03:34
Activate the production Newport resident path and explicit frontend modes. Complete IMPACT engine fidelity and snapshot lifecycle, make VPro GPU failures fail closed with full Octane snapshots, and add complete GR2 lifecycle state plus measured resident-GPU decision telemetry.

Keep VPro bit-exact on Vulkan and D3D12, and retain gr2-jit after the evidence gate does not justify a resident RE3 wgpu implementation.
@Tzeentchnet
Tzeentchnet force-pushed the feature/gpu-upstream-m9-m16 branch from 226a785 to 786a730 Compare November 17, 2026 04:01
@Tzeentchnet

Copy link
Copy Markdown
Contributor Author

Performance context

We do not currently have a committed plot, but the application-level acceptance data is sufficient to show the measured improvement over the ordinary CPU raster/JIT path.

The comparison used the same installed IRIX 6.5 IP22 system and isolated overlays for CPU raster/JIT, Vulkan resident direct, and D3D12 resident direct. Atlantis ran after a 15-second warm-up for a 30-second measured interval on an RTX 5070 host with 16 logical CPUs.

Engine REX3 GO/s Relative throughput Host core equivalents Host CPU vs CPU
CPU raster/JIT 4,761.5 1.00x 1.673 baseline
Vulkan resident direct 6,765.6 1.42x 1.574 -5.9%
D3D12 resident direct 5,918.0 1.24x 1.545 -7.7%
REX3 throughput, normalized to CPU

CPU     1.00x  ##########
Vulkan  1.42x  ##############
D3D12   1.24x  ############

Both GPU samples required unsupported=0, no_cpu_fallback=true, no increase in raw or composed readback counters during the timed interval, and conservative p95 presentation buckets no higher than 4.2 ms. The scripted Showcase crop was pixel-identical on all three paths.

This should be read as a live Newport/REX3 graphics-command throughput result, not a claim of 1.24-1.42x total emulator or guest-application FPS. IMPACT and VPro have separate workload/replay measurements, while GR2 retained gr2-jit after its resident-GPU evidence gate did not pass; combining those into one overall IRIS speedup number would be misleading.

Committed details:

@techomancer
techomancer merged commit 29be606 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