Skip to content

Development

OpenAI Codex edited this page Nov 2, 2026 · 10 revisions

Development

Reviewed: 2026-11-01 · Source: b7140848 · Availability: current main; release differences are explicit below.

Current status, release availability and the complete change register describe what exists now. Architecture pages describe mechanisms; dated PR reviews retain their own tested-head boundaries.

Delivered foundations

Hybrid#71, vertex freshness#85/#125, bounded vertex versions#89, three Vulkan cache layers, mono#182, JIT selection#183, ADPCM cache#187, mapping#190/#275, workers#200/#201/#305, format#191 and generated status#296 are present in reviewed main. Depth#282–#289 provides foundations while runtime consumers#290/#291 remain held. Implemented mechanisms explains purpose/defaults/limits; merge titles that still say HOLD do not override source presence.

Work that remains

DSP capture, replay and differential testing documents the executable experimental kit in PR #384, reviewed November 2. It includes build instructions, opt-in recording, headless production GP/EP comparisons, seeded campaigns and failure reduction. Availability is restricted to that PR branch; public corpus and qualification work remains under #383.

Fork changes worth contributing upstream is the outgoing shortlist, with per-PR summaries, measured benefits/correctness outcomes, overlaps and qualification gaps. It is separate from the existing upstream PR tracker.

Research and retired work separates GitHub state from recommendation and availability. Examples include GL recycling#319 stale traces, original precise-store reproducer#317, ARM ordering#320 native coverage, ARM hard-FPU#321 semantic repair and X3#351 qualification. Do not repeat stopped benchmark campaigns without a concrete new mechanism. Upstream shortlist preserves the September review and adds current states/head comparisons.

Evidence and contribution

Use overall game methodology, the repository's AGENTS.md, and controls. Preserve fixed baselines, exact executables/settings and failed attempts; report correctness, cadence, tails and CPU together. Raw benchmark records stay in linked evidence storage rather than product commits. Documentation-only work has runtime performance N/A.

History

The September26 snapshot remains dated history. PR lessons preserves earlier per-PR designs and links later work through the current register. Shader Browser's old merged state precedes its removal#268; closed#87 ideas partly informed delivered#89. History should explain those transitions rather than silently relabel old decisions.

Wiki maintenance defines the metadata, checks and separate publication step.

Clone this wiki locally