Repository navigation
feat(render): S5e + Extract-free — GPU frame in the unified scheduler - #10
Merged
Merged
Conversation
…#8343) Quality gate red on clippy/test: rustpython-vm 0.5.0 (latest published) does not compile against libc >= 0.2.187 — POSIX_SPAWN_SETSID on linux-gnu became c_short (i16) while nix 0.30 PosixSpawnFlags is a c_int bitflags: error[E0308]: mismatched types (expected i32, found i16) rustpython-vm-0.5.0/src/stdlib/posix.rs:1812 The outdated stage is a hard gate (all deps latest), so pinning libc back is not an option; no patched rustpython-vm release exists (0.5.0 is max_version, 2026-03-31) and main's fix rides on an unpublished refactor. Fix: vendor the published 0.5.0 sources (RustPython tag 0.5.0, 39a6486) into third_party/rustpython-vm with exactly the upstream fix from RustPython#8343 applied (.into() on POSIX_SPAWN_SETSID) — diff vs tag is that one hunk. Wired via [patch.crates-io] path source (deny.toml allows path deps, allow-git = []; cargo-outdated/audit skip path deps). Workspace exclude prevents auto-membership (clippy -D warnings / fmt / test stay off third-party code); rustqual.toml excludes third_party/** from the ratchet; Cargo.lock swaps the entry to a path source. Delete third_party/rustpython-vm + the patch section when a compatible rustpython-vm release is published (README there has the checklist). Co-authored-by: arena-agent <297053741+arena-agent@users.noreply.github.com>
|
❌ Quality gate failed on 📊 QUALITY SUMMARY📜 Tail of log (200 lines)Full logs are in the |
Full quality.log (run 34149520691) — part 1/6
|
Full quality.log (run 34149520691) — part 2/6
|
Full quality.log (run 34149520691) — part 3/6
|
Full quality.log (run 34149520691) — part 4/6
|
Full quality.log (run 34149520691) — part 5/6
|
Full quality.log (run 34149520691) — part 6/6
|
| (stale part — log got shorter) |
The S5d/S7 commit added crates/render/src/gpu_resources.rs storing wgpu Device/Queue/Surface/CommandBuffer/Renderer3D as World resources, but wgpu's web backend types are !Send/!Sync (Rc, JS callbacks) while World::insert / Resources::get / SystemAccess require Send + Sync — 26 x E0277 on cargo check -p ornis-wasm --target wasm32-unknown-unknown. Hidden on master by the 10-annotations-per-step cap (clippy/test flood from the rustpython-vm break); the PR digest summary exposed it. Native-only module: the only consumer outside ornis-render is the native root binary (src/main.rs); the wasm path renders through RenderWorld extraction + ornis-wasm, as before S7. Same convention as frame_exec.rs cfg pairs. Docs synced (README, unified-scheduler.md). Co-authored-by: arena-agent <297053741+arena-agent@users.noreply.github.com>
Author
|
✅ Quality gate is green — run 34107349433 passed all 13 stages (fmt / clippy / rustqual / smoke / test / test (physics gpu) / clippy (physics gpu) / audit / deny / outdated / doc / wasm-check; upgrade-check = SKIP, cargo-upgrade not installed in CI — pre-existing). The ❌ digest + log chunks above are from the previous (intermediate) run and are kept for history — the workflow only overwrites them on failure. |
First step of the S5e + Extract-free decomposition (docs/rendering/unified-scheduler.md): the pass-as-system adapter. - schedule_bridge: try_project_schedule(&SystemSet) mirrors every enabled pass into a core Schedule PassSystem twin — same name, accesses projected from ResourceId to the FrameResource TypeId via the registry. Disabled passes drop out with their edges, mirroring layout_levels exactly; untyped resources are an honest ProjectionError (nothing to mirror). run() is a no-op declaration twin by design — E2 moves recording into systems via FrameCommandBuffers. - RenderFrame3D::render_schedule: renders a frame ordered by the projected Schedule levels (flatten → execute_in_order), recording through the caller's borrowed encoder — same dispatch_pass as the sequential/parallel paths (extracted into PassDispatch). Debug-assert pins projected levels == FrameLayout::levels() every frame. - Gates: scheduler_parity extended with the E1 canon (projection == layout levels on chains/shared-levels/edge-splits, disabled-pass culling, untyped rejection) + production Technique×bloom matrix in bridge unit tests; tests/schedule_render.rs pins pixel identity schedule-driven vs sequential on lavapipe (harness mirrors the S5b parallel_render gate). - Docs: unified-scheduler.md E1 section marked done, README row synced. Co-authored-by: arena-agent <297053741+arena-agent@users.noreply.github.com>
|
❌ Quality gate failed on 📊 QUALITY SUMMARY📜 Tail of log (200 lines)Full logs are in the |
… harness Three findings from the E1 gate run, fixed at the source: - system.rs: resource_type used an explicitly-dereferencing pattern inside an implicit borrow (edition 2024 hard error); the find closure now compares through the double reference instead. - fmt: schedule_bridge module/use sorted to rustfmt order in lib.rs, two wrap forms in bridge/parity restored to rustfmt's preference. - rustqual ratchet (+25 findings): tests/schedule_render.rs was a near-verbatim copy of parallel_render.rs — DUPLICATE exact + a dozen FRAGMENTs. Extracted the shared headless-GPU harness into tests/common/mod.rs (scene, targets, readback, render_frame_pixels, with_headless_scene skip entry, sequential_reference_pixels); both gates are now ~30 lines of unique logic each. Removes the pre-existing duplicate/fragment findings from parallel_render.rs too, so findings land below the 1813 baseline. render_schedule is now fallible (Result<(), ProjectionError>, # Errors documented) instead of panicking — the projection failure is a declarable error, and the gate test's expect is not flagged in test code. Co-authored-by: arena-agent <297053741+arena-agent@users.noreply.github.com>
|
❌ Quality gate failed on 📊 QUALITY SUMMARY📜 Tail of log (200 lines)Full logs are in the |
….33.7 - scheduler_parity: unwrap_err() on Result<Schedule, _> requires Schedule: Debug (it isn't, and shouldn't need to be) — switch to .err().expect(), Option::expect has no Debug bound. - glam 0.33.7 released 2026-09-07 12:05 UTC; the outdated stage is a hard gate, so bump the lock entry (all workspace requirements are caret 0.33; 0.33.6 lock entry had no active optional deps, 0.33.7 same features — version+checksum swap only). Co-authored-by: arena-agent <297053741+arena-agent@users.noreply.github.com>
Recording moves off the borrowed encoder on the native runtime path: per-pass encoders land in FrameCommandBuffers (registration order), RenderPresent hands the acquired frame to FramePresentTarget, and the new RenderFlush system does the single ordered submit + present (WaW on both handover resources, registration order). - FrameExecutor::record_in_order — sequential form of execute_parallel mechanics (own encoder per pass, buffers pushed in registration order) - RenderFrame3D::render_to_buffers — second recording form over the shared projected_order helper (no submit); borrowed render() stays for wasm/examples/benches - FrameCommandBuffers::flush — drain + single ordered submit - RenderPresent: acquire → reconfigure-on-Outdated/Lost → render_to_buffers → stash into FramePresentTarget - gate: third pixel-parity test on the shared GPU harness (sequential reference vs buffers+flush) - docs: unified-scheduler.md E2 section + README row Co-authored-by: arena-agent <297053741+arena-agent@users.noreply.github.com>
First Extract-free step: upload data no longer round-trips through the Mutex<RenderExtracted> snapshot clone. RenderSubmit declares reads_lane on TransformDesc/MeshDesc/MaterialDesc (S5d canon, same shape as RenderExtract) and calls the shared extract_render_data canon on the SmartStore at submit time. Ordering vs lane writers is now lane-granular RaW instead of RaW on the whole snapshot. RenderExtract/Mutex<RenderExtracted> stay as the oracle; the gate test pins a direct extract_render_data call byte-equal to the snapshot the scheduled system published (materials as Pod bytes, instances field-wise; three material kinds, varied tessellation). docs: unified-scheduler.md X1 section + README row Co-authored-by: arena-agent <297053741+arena-agent@users.noreply.github.com>
Mesh + tessellation cache move out of GpuFrameState into a standalone Mutex<GpuMesh> resource: re-creation no longer holds the renderer/frame plan lock (documented lock order: GpuMesh before GpuFrameState). - max_mesh_params in extraction.rs: the re-create criterion as a canon alongside extract_render_data — max tessellation over COMPLETE entities only, (32, 24) floor; iterator form (lane walk in closures) - new RenderMesh system (install_render_mesh): reads the three lanes (S5d canon), writes only Mutex<GpuMesh> - RenderSubmit drops the mesh re-create (and its GpuDevice read); RenderPresent reads the mesh resource (RaW after RenderMesh), the unsafe split-borrow narrows to renderer + frame3d - gates: canon-tie unit test (incomplete 96/64 entity must not move the maximum) + tessellation probe on the shared GPU harness (tests/mesh_resource.rs: 48/32 -> +16/12 no-op -> +64/40 re-create, vertex/index counts vs a reference sphere) - record_in_order: PassViews inlined into the run() call, mirroring execute_in_order exactly (delegation-only loop body) docs: unified-scheduler.md X2 section + README row Co-authored-by: arena-agent <297053741+arena-agent@users.noreply.github.com>
RenderSubmit no longer hardcodes set_lights: ambient + directionals come from the new RenderLights resource, written by the scene loader between frames (RenderWorld::replace_scene publishes scene lighting; RenderLights::from_scene is the shared constructor) and only read inside the schedule — no Mutex, same contract as GpuSurfaceState. - RenderLights::default reproduces the legacy hardcoded rig via consts (install_gpu_resources inserts it: a runtime without a scene renders exactly as before) - set_lights_args is the single LightDesc -> (direction, intensity, color) conversion; wasm's build_gpu_scene migrates onto it (the inline match is gone) - gates: unit tests (default == legacy args; replace_scene publishes scene lighting) + pixel probe tests/light_resource.rs on the shared GPU harness — legacy hardcoded arguments vs the resource through set_lights_args must be pixel-identical docs: unified-scheduler.md X3 section + README row Co-authored-by: arena-agent <297053741+arena-agent@users.noreply.github.com>
…shot Extract-free finished: the render frame payload is read straight from the component lanes; the scheduled snapshot round-trip is gone. - RenderExtracted renamed to FrameUpload (a direct-read payload, not a snapshot); the RenderExtract system, the Mutex<RenderExtracted> resource and install_render_extract are deleted - RenderWorld::extracted() -> frame_upload() (extract_render_data on the world's store); RenderWorld stays the scene-loader host — its schedule is empty now - RenderPresent (the last snapshot reader) reads SmartStore + the three lanes (same access shape as RenderSubmit/RenderMesh); the instance count comes from extract_render_data directly - wasm: GpuScene.extracted is a FrameUpload, build_gpu_scene and the contract test use frame_upload(); schedule().len() == 0 asserted - net quality effect: one error-handling finding removed (RenderExtract ::run), none added docs: unified-scheduler.md X4 section, checklist E3 no-op noted, the decomposition header closed; README row Co-authored-by: arena-agent <297053741+arena-agent@users.noreply.github.com>
|
❌ Quality gate failed on 📊 QUALITY SUMMARY📜 Tail of log (200 lines)Full logs are in the |
CI on ac33ca0 (run 34147754661) failed fmt, clippy, rustqual and test: - fmt: apply the canonical rustfmt shapes (import ordering with cfg attributes, collapsed filter closure, split let-pattern, multiline export list, three-line assert_eq with message, exploded tuple). - clippy: drop two stale 'mut' bindings; render_to_buffers took 8 arguments — group its inputs into BufferRenderContext (native-only, published from frame_exec). - rustqual: RenderSubmit::run regressed to an IOSP Violation (logic + own call) after X1/X3 — extract the camera math into camera::camera_view_projection (pure math, no own calls) with a unit test; run() is pure orchestration again. - test: the root showcase package (src/) still used the deleted snapshot API — install_gpu_resources gains the GpuMesh resource (7 args), GpuFrameState drops mesh fields, engine_runtime tests read lanes directly via extract_render_data; the physics-ordering test is renamed to physics_sync_output_is_visible_to_render_lane_reads. - docs: X4 section notes the root-package migration; stale 'oracle stays' wording corrected. Co-authored-by: arena-agent <297053741+arena-agent@users.noreply.github.com>
|
❌ Quality gate failed on 📊 QUALITY SUMMARY📜 Tail of log (200 lines)Full logs are in the |
- clippy -D warnings: the light_resource gate binary does not use the shared sequential reference (its subject is the legacy rig, which the harness does not bake), and mesh_resource uses only the device part of the harness — annotate 'mod common' with allow(dead_code) and a comment in both, matching the shared-harness-subset reality. - fmt: split the world.engine_mut().world_mut().insert chain in mesh_resource (chain width). Co-authored-by: arena-agent <297053741+arena-agent@users.noreply.github.com>
Co-authored-by: arena-agent <297053741+arena-agent@users.noreply.github.com>
…rics Green-run summary (CI 34150634051 lineage, 13/13 PASS): - quality_score 15.9% -> 17.9% (0.17868384215287147) - iosp_score 92.2% -> 92.4% (0.9235820895522389) - total_findings 1813 -> 1807, violations 128 (unchanged) - functions 2587 -> 2640; violation_details re-pinned to current lines (128 entries). Values reconstructed from the run summary (all seven dimension percentages match the printed block bit-for-bit; scores recomputed with the exact rustqual formulas and rustqual.toml weights). Co-authored-by: arena-agent <297053741+arena-agent@users.noreply.github.com>
All five review questions are answered by the shipped implementation: module-inside-render confirmed, immediate per-frame plan confirmed, pool keyed by the full TextureSpec, sub-pass deferred (no wgpu input-attachments), debug dump delivered as the shared S6 mermaid() projector instead of a bespoke JSON dump. Co-authored-by: arena-agent <297053741+arena-agent@users.noreply.github.com>
arena-ai-coding-agent
Bot
changed the title
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes the final phase of the Unified Scheduler roadmap (
docs/rendering/unified-scheduler.md): the S5e + Extract-free decomposition.Main part — S5e + Extract-free (E1–E2, X1–X4, E3)
Schedulesystems:schedule_bridgemirrors every enabled pass into a typed-access twin (disabled passes drop with their edges);render_schedulerenders by the projected levels; level parity is debug-asserted every frame.FrameCommandBuffersin registration order, one ordered submit+present inRenderFlush;render_to_buffers(BufferRenderContext).RenderSubmitreads theTransformDesc/MeshDesc/MaterialDesclanes directly through the sharedextract_render_datacanon — no moreMutex<RenderExtracted>clone.GpuMeshresource +RenderMeshsystem: the sphere is re-created whenmax_mesh_paramsmoves (floor 32×24);GpuFrameStatekeeps only renderer + frame3d.RenderLightsresource (ambient + directionals, published by the scene loader; the default reproduces the legacy hardcoded rig bit-for-bit).RenderExtracted→FrameUpload: the snapshot system and resource are deleted; wasm and the root showcase migrated to direct lane reads; the render side now registers zero scheduled systems.Gates: ~15 new (pixel parity ×3, canon-tie, tessellation probe, light probe, scheduler parity, unit tests). CI Quality 13/13 green; ratchet improved — quality 15.9→17.9%, IOSP 92.2→92.4%, findings 1813→1807, violations 128 (flat). Baseline re-pinned to the new metrics.
Prerequisite — unblocking the Quality gate
rustpython-vm 0.5.0+ the upstream libc fix (RustPython#8343): the published crate does not compile against libc ≥ 0.2.187 (POSIX_SPAWN_SETSIDbecamec_short), and theoutdatedstage forbids pinning libc back.gpu_resourcesoff on wasm32 (wasm-check): wgpu web types are!Send/!Sync.