Skip to content

Bring Preset K back as a second model choice #15

Open

Description

@phantoslayer21

Preset K's implementation is in clean/k (reference, kernels, runtime layer and tests) and its converter map is
data/preset-K.v1.map.json, but 0.6.0 is built without it: -DWG_MODEL_K=OFF for the front end, --no-model-k for
the runtime, -DWRONGGPU_PRESET_K=OFF for the launcher (BUILDING.md).

To do: build with Preset K turned on, run the CPU tests of clean/k/tests (the modules that need a GPU are listed in
BUILDING.md, section 6), fix what breaks, and test in games on a Radeon RX 9000-series card. A pull request that turns
the options back on says what was tested and where.

Activity

  1. added this to the 0.6.1 milestone on Nov 13, 2026
  2. modified the milestones: 0.6.1, , 0.7.0 on Nov 13, 2026
  3. j0nsh1n commented on Nov 17, 2026

    @j0nsh1n
    Contributor

    Some first in-game numbers for Model K, for whatever they're worth. Please read the caveats first: none of this is a release build or a supported setup.

    Caveats

    • Linux, Cyberpunk 2077 under Proton (GE-Proton11-11), not Windows. One game, one machine.
    • Developer builds of the runtime from local branches (the RDNA 3 work behind RDNA 3 (RX 7900 XT): Model M in HIP on gfx1100 and a developer-only D3D12 route under Proton (draft) #259 plus a frame-capture diagnostic), not 0.6.x as released. On the RX 9070 XT none of the RDNA 3 code is active, so it runs main's RDNA 4 Model K kernels, but it is still a developer DLL.
    • The installed front end is the 0.6.0 release, which doesn't know Model K; K was selected with WG_RUNTIME_OPTIONS=model=K (the overlay still says Model M).
    • On the RX 7900 XT, K runs through the unofficial RDNA 3 route: a locally patched vkd3d-proton/dxil-spirv build and native gfx11 forms of some stages. The Model K: softmax weight rounding in one v_fma_mix*_f16 (exact), with a probe and a logged fallback #254 softmax self-check refuses gfx11, so those runs used the diagnostics-only k_smx_test=mix_unchecked hook.
    • The model is converted locally from the game's own nvngx_dlss.dll (310.9.1.0, preset K; all 12 stage digests matched). Nothing of it leaves the machine.
    • Times are the runtime's own GPU timestamps (timing_log=1;timing_stages=1), median of the 300-frame windows of one session each; 1706x960 -> 2560x1440 (1440p Quality), max settings, no RT, no Ray Reconstruction. Not a controlled benchmark.

    Numbers (GPU time per frame, median)

    Card Path Total input network output
    RX 9070 XT main's RDNA 4 kernels, self-check passed (smx=mix) 1.62 ms 0.32 0.76 0.53
    RX 7900 XT RDNA 3 route, exact 2.07 ms 0.44 1.11 0.50
    RX 7900 XT same + k_approx=a1 1.76 ms 0.21 1.08 0.48

    Image quality

    • On both cards K shows flicker on thin bright detail (lit floor grates, neon edges), some ghosting on moving thin objects, and occasional single-pixel specks. It looks the same on both cards.
    • I don't think it is a WrongGPU porting bug. Offline, with captured real game frames replayed through the runtime, the 7900 XT output matched the in-game output byte for byte. Your NumPy reference of K, run from the captured history and inputs, reproduced the same specks at the same pixels (e.g. 27,986 vs 27,987 one-frame specks in one 30-frame capture), and every stage stayed inside spec 49's limits. So it looks like K's own response to this game's inputs.
    • k_approx=a1 saved about 0.3 ms on the 7900 XT, but in game it looked clearly worse to me, although a replay of 30 captured frames shows only a small per-frame difference (it sits on thin bright lines and grows over the frames through the history). Worth a look on your side before recommending it.

    Happy to share the capture tooling or rerun anything specific.

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

    Projects

    • Status
      Todo

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions