Skip to content

Model L in games: the D3D12 executor and the L model path on every executor - #152

Merged
phantoslayer21 merged 3 commits into
mainfrom
model-l/executor-and-path
Nov 14, 2026
Merged

phantoslayer21 merged 3 commits into
mainfrom
model-l/executor-and-path

Conversation

@phantoslayer21

Copy link
Copy Markdown
Owner

Refs #89

Two faults that kept Model L (model=L) from running in a game on native D3D12, found in the first Cyberpunk 2077 runs
with #140:

  1. Executor choice (interop/backend.cpp): the executor request set model_k for every model other than M, so
    model=L took the "model K has no D3D12 port; the HIP executor runs it" route. Now only model K takes that route;
    Model L goes to the D3D12 executor (the native D3D12 frame of Model L frame on native D3D12: Model M's native frame with Model L's encoder and decoder #140). Model M and model K are unchanged.
  2. Model file path (interop/backend.cpp configure_hip, the configure step both executors share): with model=L
    it still required Model M's file (<model_dir>\model.wgcm) and refused with "no converted model at ...". New
    model_wgcm_path in model_select.h gives each model's file from the model table: Model M
    <model_dir>\model.wgcm (unchanged), Model L <model_dir>\L\model.wgcm, used by the configure step.

Test: clean/runtime/interop/tests/test_model_path.py (CPU, run by the Python test set) builds
model_path_check.cpp and checks model=L -> <model_dir>\L\model.wgcm, model=M unchanged, model=K's folder, and that
the option model=L resolves to Model L.

Tested:

  • The new CPU test passes locally (Ran 1 test ... OK).
  • In game, on an integration build with these commits (model-l/ingame adea429, with Model L frame on native D3D12: Model M's native frame with Model L's encoder and decoder #140 and the Model L build
    switch): Cyberpunk 2077 and Resident Evil Requiem ran Model L on the D3D12 executor at 1280x720 -> 1920x1080. The
    runtime's launch trace showed every Model L kernel (lev/input_encoder_direct, l_swin_block_c32_enc_nc /
    _dec_nc, l_coef_pack12, ...) once per frame, with 0 failed dispatches. Records and screenshots are kept with the
    release lead's run notes.

Not in this PR: the Model L build switch (--model-l) and the launch_trace runtime option used for that evidence;
they follow separately.

… Model L goes to the D3D12 executor

Signed-off-by: phantoslayer21 <156617979+phantoslayer21@users.noreply.github.com>
…ared configure checked Model M's)

Signed-off-by: phantoslayer21 <156617979+phantoslayer21@users.noreply.github.com>
…y executor); CPU test test_model_path.py

Signed-off-by: phantoslayer21 <156617979+phantoslayer21@users.noreply.github.com>
@phantoslayer21
phantoslayer21 marked this pull request as ready for review November 14, 2026 07:30
@phantoslayer21
phantoslayer21 merged commit e53723b into main Nov 14, 2026
9 checks passed
@phantoslayer21
phantoslayer21 deleted the model-l/executor-and-path branch November 14, 2026 07:49
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