Skip to content

Build: the Model L switch (--model-l, off by default) and the D3D12 kernel sets - #154

Merged
phantoslayer21 merged 5 commits into
mainfrom
model-l/build-switch
Nov 14, 2026
Merged

phantoslayer21 merged 5 commits into
mainfrom
model-l/build-switch

Conversation

@phantoslayer21

Copy link
Copy Markdown
Owner

Refs #89

The Model L build switch used for the first game runs of Model L (Cyberpunk 2077, Resident Evil Requiem), off by
default so that a build without it is Model M's build.

  • clean/runner/build.py --model-l: Model L's runtime sources (clean/l/runtime) compiled with WG_MODEL_L, its code
    objects (l_swin, l_ops, l_post, l_enc_values, built as clean/l/runner/l_c32_check.py builds them) and
    preset-L.v1.map.json built into the runtime library. With --runtime-only --rocm the code objects are built first
    for every target (release builds reuse pinned Model M code objects).
  • clean/dx12/build_dx12.py --model-l: Model L's D3D12 kernels (every port of an l_*.co, and dec1's standalone block
    swin_block_c64_add_ex_nc, which only Model L launches) in the manifest; without it they are left out unless
    --only names them (the Model L checks use --only). --list prints the kernel set of a build, compiling nothing.
  • clean/dx12/kernel_sets.json + clean/dx12/tests/test_kernel_sets.py (CPU): the declared sets, Model M's (46) and
    what --model-l adds (14), checked in seconds before any build; a kernel added or moved updates the list in the
    same change (python clean/dx12/tests/test_kernel_sets.py --write).

Model M when L is off: build.py builds and embeds exactly what it did. build_dx12.py compiles the same Model M
kernels; the default manifest no longer carries Model L's 14 kernels (main's default build currently includes them;
no release ever shipped them).

Tested: the kernel-set test passes locally; an integration build with the switch on (runtime, front end, launcher,
package) passed its stages and ran Model L in both games on the D3D12 executor.

…ernel sets

clean/runner/build.py --model-l: Model L's runtime sources with WG_MODEL_L, its code objects (l_swin, l_ops, l_post,
l_enc_values, as l_c32_check.compile_kernels builds them) and preset-L.v1.map.json built into the runtime library;
with --runtime-only and --rocm its code objects are built first for every target.

clean/dx12/build_dx12.py --model-l: Model L's D3D12 kernels (co=l_*.co, and dec1's standalone block, which only Model
L launches) in the manifest; without it they are left out unless --only names them. --list prints the kernel set
of a build without compiling anything.

kernel_sets.json and tests/test_kernel_sets.py: the declared sets (Model M's, and what --model-l adds), checked in
seconds before any build. Without the option both builds are Model M's: the same Model M kernels; the default D3D12
manifest no longer carries Model L's kernels.

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 08:03
…he Model L set

l_linear_32_32_fused (#160) and l_swin_block_c32_dec_coef12_nc (#159) have no
HIP code object, so the co test alone put them in Model M's set. Model M's set
is unchanged (46); Model L's set is 16. kernel_sets.json regenerated; the
profile tool's docstring names --model-l.

Refs #89

Signed-off-by: phantoslayer21 <156617979+phantoslayer21@users.noreply.github.com>
@phantoslayer21
phantoslayer21 merged commit 9a42f4b into main Nov 14, 2026
9 checks passed
@phantoslayer21
phantoslayer21 deleted the model-l/build-switch branch November 14, 2026 11:44
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