Fix: cudnn-sys Windows build include paths

[CharryWu]

Summary

• Extend cudnn-sys Windows cuDNN detection to support modern CUDA Toolkit include paths under C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\….
• Add dynamic discovery of vX.Y directories (e.g. v10.6, v12.1, v13.0) under both the legacy cuDNN root and the CUDA Toolkit root.
• Add unit tests that validate the vX.Y directory-name matching logic and that multiple Windows CUDA/cuDNN versions are discovered correctly.

This change is related to my earlier comment on issue #204, where cuDNN headers could not be found when using newer CUDA/cuDNN layouts.

Details

• Windows include path discovery

  • Previously, Windows builds only checked:
    • C:/Program Files/NVIDIA/CUDNN/v9.x/include
    • C:/Program Files/NVIDIA/CUDNN/v8.x/include
  • Now, for Windows:
    • We keep the above legacy paths as explicit fallbacks.
    • We dynamically enumerate directories under:
      • C:/Program Files/NVIDIA/CUDNN
      • C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA
    • For each of these bases, we treat any subdirectory whose name matches vX.Y[.Z…] (all numeric components) as a candidate version and add <base>/vX.Y/include to the search list.
    • The first directory that contains both cudnn.h and cudnn_version.h is selected.

• Helper functions

  • Introduced CudnnSdk::is_vxy_dir_name to encapsulate the vX.Y pattern check.
  • Introduced CudnnSdk::collect_windows_cudnn_include_paths to gather include paths from one or more base directories.

• Tests

  • is_vxy_dir_name_accepts_valid_versions and is_vxy_dir_name_rejects_invalid_versions ensure the matcher behaves correctly for valid/invalid names.
  • collect_windows_cudnn_include_paths_discovers_multiple_versions creates a temporary directory structure with:
    • v10.6/include, v12.1/include, v13.0/include (with stub cudnn.h and cudnn_version.h),
    • plus some non-version directories,
      and asserts that only the valid vX.Y/include directories are discovered.

Rationale

• Newer CUDA Toolkit + cuDNN combinations commonly install under C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.0\include and similar, which the previous hard-coded paths did not cover.
• Allowing discovery of any vX.Y directory under the known NVIDIA roots makes cudnn-sys more robust across:
  • different major/minor CUDA versions,
  • users who keep multiple toolkits installed side-by-side,
  • future versions where only the numeric components change.

Testing

• cargo check -p cudnn-sys on Windows with:
  • CUDNN_INCLUDE_DIR unset,
  • real v13.0 CUDA Toolkit install,
  • additional copied directories v12.1, v10.6 under C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA.
• cargo test -p cudnn-sys (Windows):
  • All new unit tests in crates/cudnn-sys/build/cudnn_sdk.rs pass.

Attribution

This bugfix was implemented with the help of Cursor.