Skip to content

0.6.6: follow-ups from the Model L hybrid default review (#256) #258

Open

Description

@phantoslayer21

Follow-ups from the review rounds of #256 (Model L hybrid form as the native D3D12 default). None of them changes what a default run outputs. They are tracked here for 0.6.6.

1. Wider game-path timeout test

The forced-timeout test through the game interface covers a timeout on a later frame, one output format on the direct route, and RGB values. Extend it to:

  • both proxy routes (sRGB or non-UAV outputs as well as UAV outputs);
  • byte comparisons that include alpha;
  • an exact count of every log line emitted for the stuck, refused and reset frames;
  • the reset compared with a fresh context's first frame in the pipelined mode as well as the synchronous one.

2. Legacy partial manifests

A D3D12 kernel manifest written before probe entries were marked can keep an unmarked probe (for example l_q7_s2a_probe_m1) through a later partial --only build, and the embedding step would accept it. Resolve manifest entries against the current kernel declarations, or refuse unknown and legacy entries fail-closed, and extend the release diagnostics check to that probe's file name. Add a legacy-manifest negative control.

3. ISA output of scan mutants

With --isa, the scan mutants' ISA and statistics are written under the build's dx12/isa folder, outside the temporary folder that holds the mutants. Keep mutant diagnostics in the temporary folder, and test the --isa build as well as the default one.

4. FP16 representability of scalar literals in the ISA rule

The conversion rule accepts any literal moved into a scalar register as an FP16 value. A literal that is FP32 but not FP16-exact (for example one just above an FP8 midpoint) would pass, although rounding it directly to FP8 can differ from rounding it through FP16 first. Track literal bit values and accept them as conversion sources only when they are FP16-exact. Add that negative control.

5. Barrier and kernel-substitution switches on the Model M route

The barrier and kernel-substitution switches are ignored on the release Model L hybrid route but are still honoured on the release Model M route. Limit them to diagnostic builds on every route, and extend the release diagnostics check to them.

Activity

  1. phantoslayer21 commented on Nov 16, 2026

    @phantoslayer21
    OwnerAuthor

    More follow-ups from the last review round of #256, tracked here rather than in 0.6.5. None of them changes a normal frame's output.

    6. Refused frames without a stand-in

    With no stand-in available, a run of "this frame only" refusals (WG_ERROR_NOT_READY) is counted by the front end's existing bad-frame limit, and the 30th in a row hands the context to the alternative upscaler permanently. That only happens when the game's predicate can never be reported, where Model L hybrid cannot run anyway. Even so, refusals of this kind should not count towards the bad-frame limit, so that a later frame with a known predicate can still reach the network.

    7. Untracked command lists and the split check

    A command list whose vtable the front end cannot track is rejected by the existing split check, and that rejection goes through the permanent handoff. This behaviour predates #256. Consider routing it through the one-frame refusal instead.

    8. Stand-in under a suppressing game predicate

    The stand-in records its dispatch without lifting a game predicate that suppresses it, and then reports success. On a refused frame, the output would be left unwritten. Lift and restore the game's predicate around the stand-in's work, as the network path now does.

    9. Stand-in on an output without unordered access

    The stand-in creates a UAV for the game's output even when that texture lacks unordered access, which a valid RGBA16F copy-route output can. Check for unordered access before using the stand-in. If the texture lacks it, write through a proxy texture or hand that one frame to the alternative upscaler.

    10. Real-data guards for older or copied cases

    • A real cases file made before the provenance markers existed, or copied without its marker, is treated as synthetic. Each run path should also check the real-data flag embedded in the file, or refuse unmarked cases in a repository folder before reading them.
    • A synthetic stage-check run can read dump and output files that a real run left in its folder. Check provenance on that path too.
    • The private JSON check resolves only the parent folder, so a symlinked output file can point back into the repository. Resolve the full output path.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions