Repository navigation
Publish GPU grid research with independent CUDA and Vulkan approaches - #150
Conversation
|
Important
This repository does not receive automatic reviews because it has fewer than 10 stars. ⚙️ Run configuration
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
|
Cleanup disposition: retain draft at exact head |
|
PR142 serial integration lead active; base |
Fresh source integration review — blocked on final artifact accessPR150 remains OPEN / DRAFT, not merge-ready, at exact remote head Source and independent reviewPrepared a clean local no-commit normal merge with that main: exact composed tree Independent Pandoc parsing reconfirms all 25 original mathematical expressions retained among 42 current expressions, zero chapter CodeBlocks, 176 relative file links and zero missing file/Markdown fragment targets. Owner independently checked the same 176 file destinations; rendered and public-link qualification is not inferred. The historical CPU pin, dated source refresh, primary-source partial-access boundaries, exact BSD comparisons and prohibition on importing unreviewed upstream expression/code remain explicit. Independent selective primary-source checks supported the CUDA/Vulkan ownership/synchronization and scoped literature comparisons; this is not an exhaustive new upstream audit. Workflow normalization reproduces main after removing comments and reversing only the previously approved 120-minute cap to 90. SHA256 Fresh source-only supporting checks: staged/unstaged whitespace passed; Exact-head CI and artifact gateExisting exact-head normal attempt-1 workflows reconfirmed successful: full book 39643500126, compressible preview 39643500346, dense3D metadata 39643500195. Full build and bounded lessons succeeded; PR deployment was skipped. These retained results belong to Read-only metadata reconfirms final artifact BLOCKED — exact action/target/reason: download and inspect final artifact Minimal recovery: obtain materially new explicit authorization/access for one direct supported download of that exact existing final artifact; verify its archive digest, bound source/book/PDF/browser receipts, GPU HTML/Markdown/source links, and all final GPU PDF pages. Do not rebuild or rerun sealed campaigns to replace the denied artifact. The recovered original-head artifact would still not qualify the thermal-composed tree: before later normal merge, the lead must revalidate the actual integration head against then-current main and obtain its required matching publication evidence. Coordination and handoffNo serial merge is authorized. PR142 lead owns the safe integration point; this worker will not merge without a verified exact-head authorization comment and all required gates satisfied. Source composition is reviewed against main15026a9 only; later product/native integration may require fresh reconciliation. PR150 supplies research exposition and proposed independent CUDA/Vulkan paths; it supplies no GPU solver implementation, supported device/hardware qualification, CPU/GPU parity, restart equivalence, convergence or measured speedup. This PR cannot establish whole-application production acceptance. Local resource retained-protected: Reconfirmed GPU source SHA256 identities
|
Current integration status (2026-10-21): BLOCKED / retain draft, exact head
2da749bd08b7de44990b345b3b807736f18b8487. All three required exact-head CI workflows passed. Fresh independent source review of clean thermal-main composition treeffa2e55ed941530b783f0fbf8e5cda8fafe9125ffound no deciding source defect; 56 source/fixture test executions passed. Final artifact12369884006content download was previously denied (Forbidden) and has not been retried or substituted. Required final content/provenance/PDF visual review remains unsatisfied. Current review, source hashes, results and exact recovery/serial-integration conditions. PR142 lead owns merge authorization. Historical preparation/run text below retains its original context; its then-running workflows are now successful, without waiving the artifact gate.Publish Chapters 37–39 on grid-fluid GPU algorithms, independent CUDA and Vulkan compute paths, and CPU-reference validation/residency/implementation gates. Include the CPU inventory, primary-source bibliographies and source/license/access audit, with one additive book README discovery paragraph. Puma remains the credited author.
Head
2da749bd08b7de44990b345b3b807736f18b8487, treeed1728a12a9e4a734dfc9d123b4df9a6150c304a, preserves research draft295e9be and publication db6f33d in normal ancestry, based on accepted main d67aec1 including portal PR148. The incremental budget commit changes exactly six authorized paths: the publication workflow cap/comment and five owned process records. GPU chapter/guide/bibliography and README hashes are unchanged. FEM30–36, portal40, thermal PR149 paths, production physics, dependencies, proofs and builders remain untouched. No force push, retarget, branch deletion or merge.The no-pseudocode/source-license rule is preserved: nine former code/diagram blocks became original prose and explicit recurrences; all25 original equations remain, with42 current expressions and no CodeBlocks. Exact BSD-3-Clause AmgX/Flow comparisons are distinguished from scientific/API factual provenance; no upstream code, figure or dependency is incorporated. Partial-access and selected-section limits remain explicit. The historical CPU oracle is fd7c0f0, with a source-only d67aec1 authored-force/recording refresh rather than relabelled numerical evidence.
The full publication run 39626898860 was cancelled after90m29s during final rejection probes. Its render, site and explicit retained browser checks passed; remaining runtime/aligned-strain suites and the final full-site artifact did not complete. Three completed full-browser traversals took20–21minutes. The configured90-minute cap strongly supports timeout exhaustion, but the detailed cancellation annotation was unavailable. The parent-coordinated remedy changes only this job’s wall-clock cap to120minutes and its explanatory comment. Every command, numerical threshold, per-case limit, sandbox setting and test is unchanged. If this one bounded trial times out, stop and re-plan rather than automatically increase it again.
Independent review passed exact head2da749b, its workflow candidate hash, all eight GPU book hashes and the README hash, with no deciding finding. Diagnosis and independent budget review retain the cancelled outcome, timing evidence and access limitations. Focused whitespace/exact-replacement/source-identity checks passed; no local science or browser workload was run for this budget-only change.
Normal new-head CI is running at attempt1 for each listed workflow, with no manual duplicate rerun:
Prior db6f33d preview 39626898869 passed actual Chromium render/check and site/package verification. Its artifact12357433282 PDF was independently visually reviewed across all GPU pages161–192, with no blocking math/table/clipping/attribution defect. PDF SHA256
5bb728859fab3bb25dacc8591d52a0aff2f93168f26471dda84a9836179edc97. That evidence retains its original source identity and does not substitute for the new workflow’s required terminal results/artifact.Keep draft until required new-head gates pass and parent artifact/integration verification completes. Parent coordinates later normal integration after thermal approval and final lead merge. This PR claims no GPU implementation, hardware support, CPU/GPU numerical parity or measured speedup. Earlier uncertain Library upload artifacts remain preserved without retry.