Skip to content

0.6.6: follow-ups from the k_approx review (#257) #262

Open

Description

@phantoslayer21

Follow-ups from the review rounds of #257 (Model K: the k_approx opt-in close approximations a1 and a2). They are tracked here for 0.6.6; none of them changes the default (nothing set) output.

1. Preparation of a selected form with an unsupported root signature

An otherwise valid manifest can declare a selected approximate form (for example k_out_pixel_c__a2) with a root signature the native device does not support. The loader accepts it, the create-time preflight checks only the name and code object, and preparation returns success without loading the shader, so create succeeds and the first launch fails. Make preparation fail for an unsupported signature or an unmatched variant, and test that malformed selected entry. (The same preparation gap is item 2 of #261.)

2. Non-vacuous selection tests

The approximation test never asserts the k_out_color__a2 launch: replacing only that selection with the exact colour kernel would still pass. The executor matrix reads back no output bytes. Assert every expected approximate launch and the absence of its exact counterpart, for buffer and direct frames and for each tag on its own. Compare direct-route frames byte for byte where an exact reference applies.

3. Mode line with a very long options variable

If the runtime options environment variable exceeds the interop parser's 4095-character limit, the front end reports that it ignored the value, while Model K reads the full variable independently and enables the approximation. The mode line can then say "exact". Derive the mode line from the runtime's effective selection, or use one options source, and test the long-value case.

4. Claims without committed records

A comment in clean/k/dx12/hlsl/k_out.hlsli states a one-ULP maximum and a difference rate for six seeds, and the PR description cites a shader-identity record and an executor negative-control record. Neither record is in the repository. Commit constructed-data summaries for these (synthetic inputs only), or qualify the claims.

5. Frame-note threading contract

wg_rt_take_frame_note hands over the same per-context note a frame can write, without synchronization. The executor calls it synchronously after a frame, which is safe, but the contract is not written down. Document that frame calls and note calls on one context must not run concurrently, or synchronize them.

6. The k_fill test option in release builds

The k_fill option fills Model K buffers with a test pattern and is accepted in release builds; it predates #257. Limit it to developer builds, as the other test hooks are, and add it to the release diagnostics check.

Activity

  1. added this to the 0.6.6 milestone on Nov 16, 2026
  2. phantoslayer21 commented on Nov 16, 2026

    @phantoslayer21
    OwnerAuthor

    Two small items from the review of #264 (the Model K direct-route legacy-input fix):

    7. Direct-legacy test assertions

    clean/k/tests/test_k_direct_legacy.py compares the direct route with the exact reference over six frames. It should also assert which executor produced each side and require a positive expected byte count, so a misconfigured run cannot pass by comparing nothing. It should also cover a single context that switches between tiled and legacy frames, so that output after approximate tiled frames have changed the history is checked too.

    8. Stale preflight comment

    The preflight comment at clean/k/runtime/runtime_k.h:66-69 still says legacy-input frames are refused; since #264 they are allocated and run. Update the comment.

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

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions