[ViralBShah]

Adds opt-in CUDA support: after ] add OpenCV_CUDA_jll and a restart, OpenCV loads its binaries from OpenCV_CUDA_jll (CUDA 13 + cuDNN, Linux x86_64/aarch64) instead of OpenCV_jll. This enables cv.dnn's DNN_BACKEND_CUDA. Removing OpenCV_CUDA_jll switches back to the CPU build.

Requires JuliaPackaging/Yggdrasil#15546 (OpenCV_CUDA_jll 4.14.0+1) for the CUDA path; see below.

Design

• OpenCV_CUDA_jll is a weak dependency, not a regular one. Pkg downloads the CUDA runtime and cuDNN when OpenCV_CUDA_jll is installed, whether or not it is loaded. The artifacts are lazy, but the .pkg/select_artifacts.jl hook of CUDA-augmented JLLs selects them with include_lazy=true. In a fresh depot on a GPU machine, a hard dependency made ] add OpenCV download about 1.8 GB: CUDA_Runtime 1023 MiB, CUDNN 431 MiB, CUDA_Compiler 188 MiB, CUDA_Driver 108 MiB and OpenCV_CUDA 42 MiB. As a weak dependency, it adds nothing to the install. The [compat] entry still constrains its version when a user adds it.
• The library is chosen in __init__. If OpenCV_CUDA_jll is in the environment and has binaries for the host, OpenCV loads it and points @wrapmodule's library callback at its libopencv_julia before @initcxx. The wrappers stay precompiled against OpenCV_jll, which works because both libraries are built from this repo's gen/. They register the same 859 bindings and export the same 25,836 symbols, and @initcxx re-resolves every function pointer by key. While output is being generated for precompilation, OpenCV always uses OpenCV_jll.
• No package extension. An extension can't do this: @initcxx binds the library in OpenCV.__init__, and extensions load only after their parent's __init__. There is also no preference to set: installing the package is the opt-in.
• Coexistence with OpenCV_jll. OpenCV_jll is still loaded, because it is a regular dependency. [OpenCV_CUDA] Suffix library names with _cuda JuliaPackaging/Yggdrasil#15546 renames the CUDA build's libraries (libopencv_core_cuda.so.414, …) so the two builds don't clash by soname. With the currently registered OpenCV_CUDA_jll 4.14.0+0, which lacks the renaming, OpenCV warns and stays on the CPU build.

Testing

Julia 1.13.3, 2× H100 NVL (driver 595.45, CUDA 13.2), OpenCV_jll 4.14.0+0:

• Fresh depot, ] add OpenCV: no CUDA artifacts are downloaded (455 MB of artifacts in total, against 4.0 GB with a hard dependency), and OpenCV binds OpenCV_jll.

• With OpenCV_CUDA_jll from Yggdrasil#15546's CI added: 313/313 tests pass, OpenCV binds OpenCV_CUDA_jll, and both builds' libraries are loaded side by side.

• With the registered OpenCV_CUDA_jll 4.14.0+0 added: OpenCV warns, binds OpenCV_jll, and 313/313 tests pass.

• DNN face detector (opencv_face_detector_uint8.pb, 300×300), with identical detections on all backends:

  Backend / target	ms/frame
  DNN_BACKEND_OPENCV / DNN_TARGET_CPU	15.1
  DNN_BACKEND_CUDA / DNN_TARGET_CUDA	3.1
  DNN_BACKEND_CUDA / DNN_TARGET_CUDA_FP16	3.0

  Model has no generated setPreferableBackend binding, so the backend is set on a Net from readNet before wrapping it in DetectionModel.

Not covered: CI here has no GPU and does not install OpenCV_CUDA_jll, so it only exercises the CPU path. I could not test a Linux machine without an NVIDIA driver, where OpenCV_CUDA_jll should have no binaries and OpenCV should stay on the CPU build.

Bumps the version to 4.11.0, a minor release because this only adds functionality. The release also includes the unreleased OpenCV 4.14 wrappers from #103.

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 37.50000% with 10 lines in your changes missing coverage. Please review.
✅ Project coverage is 89.28%. Comparing base (3263f0b) to head (1492444).

Files with missing lines	Patch %	Lines
src/core.jl	37.50%	10 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #104      +/-   ##
==========================================
- Coverage   92.79%   89.28%   -3.52%     
==========================================
  Files          11       11              
  Lines         236      252      +16     
==========================================
+ Hits          219      225       +6     
- Misses         17       27      +10     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[ViralBShah]

@yakir12 FYI.