Skip to content

Native direct check for Model L; the native candidate names the model it runs - #151

Merged
phantoslayer21 merged 1 commit into
mainfrom
model-l/direct-check
Nov 14, 2026
Merged

phantoslayer21 merged 1 commit into
mainfrom
model-l/direct-check

Conversation

@phantoslayer21

Copy link
Copy Markdown
Owner

The native direct-resource check for Model L, and a provider label that names the model it runs.

What changes

  • clean/dx12/tests/native_direct_build.py --model-l: builds the test program with Model L (WG_MODEL_L, the
    sources in clean/l/runtime).
  • clean/dx12/tests/native_direct_test.cpp: an optional last argument, the model.
    • M (the default) is unchanged.
    • L runs with the runtime option model=L and reads model.wgcm, with no output_filter=original, since Model
      L's file has no output-filter weights. Model L's map is found as preset-L.v1.map.json in the build folder.
    • OPTIONS prints the model.
  • clean/runtime/interop/native_candidate.cpp: provider() says "Model L native D3D12, ..." when the runtime runs
    Model L. Model M's text is unchanged.

Refs #89

GPU record (this branch's test program; build = main's D3D12 kernels and HIP code objects with l_frame_check.py --prepare's synthetic models; locked GPU slot, gpu_slot.py --cap-s 170)

The command:

native_direct_test.exe <build> <model> direct check 853 481 1279 721 2 1 3 3 600 1 scale 0 L
  • PASS native direct check queue=direct frames_compared=2 format_combinations=1 cycles=1 mutation=1, 2.1 s in the
    slot:
    • the direct-texture route (the game's) and the staged route give byte-identical outputs on both frames;
    • the check's own mutation is caught;
    • the D3D12 debug layer was on.
  • PROVIDERS direct=Model L native D3D12, supplied device/queue, direct textures staged=Model L native D3D12, ....
  • Launch trace (WG_LAUNCH_TRACE):
    • both routes run l_linear_32_32_relu / _f16, l_swin_block_c32_enc_nc / _dec_nc, l_project_gather,
      l_dec_project and l_coef_pack12;
    • the direct route runs lev/input_encoder_direct and w64/post_tf_f32_p12_diet_vsit_direct;
    • the staged route runs lev/input_encoder and w64/post_tf_f32_p12_diet_vsit.
  • An earlier run of the same check with the previous label (main's native_candidate.cpp) passed 4 frames, 5.5 s in
    the slot. A later 2-frame run hit the slot's 170 s cap while creating pipeline states, with the game run and a
    build on the machine at the same time; re-run idle, it passed (the record above).

Tested

  • The GPU record above, and the public gates.

Not tested

  • Model M with this program since the change: its path is the same code; its label text is unchanged.
  • In a game: that is the release lead's run (WG_LAUNCH_TRACE for the proof).

… it runs

- clean/dx12/tests/native_direct_build.py --model-l: the test program with Model L
  (WG_MODEL_L, clean/l/runtime).
- clean/dx12/tests/native_direct_test.cpp: an optional last argument, the model
  (M by default; L: runtime option model=L, model.wgcm, no output_filter=original,
  since Model L's file has no output-filter weights). OPTIONS prints it.
- clean/runtime/interop/native_candidate.cpp: provider() says "Model L native
  D3D12, ..." when the runtime runs Model L; Model M's text is unchanged.

Refs #89

Signed-off-by: phantoslayer21 <156617979+phantoslayer21@users.noreply.github.com>
@phantoslayer21
phantoslayer21 marked this pull request as ready for review November 14, 2026 06:50
@phantoslayer21
phantoslayer21 merged commit 7bc81bb into main Nov 14, 2026
9 checks passed
@phantoslayer21
phantoslayer21 deleted the model-l/direct-check branch November 14, 2026 07:50
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant