adaption to CUDA13/GCC13/ARM arch

[pcastelovigo]

Start to adapt code to compile with ARM/CUDA13/GCC13. The goal is to support NVIDIA DGX Spark workstation
I'll try to send all patches in next two weeks without breaking anything. I have it compiled, but it was arduous.

• Add CUDA 13 compute capabilities
• Remove deprecated clockRate call if using CUDA 13
• Add parameter to specify CUDA CC manually if needed
• GCC 13 requires explicit include in draco, added conditionally

[Copilot]

Pull request overview

Adapts the build system and GPU reporting code to better support building on newer toolchains (CUDA 13 / GCC 13) and ARM targets (e.g., NVIDIA DGX Spark), primarily by extending CUDA architecture/CC handling and working around a GCC 13 dependency compile issue.

Changes:

• Adds superbuild cache parameters for controlling CUDA architectures for dependencies and CUDA CC list for AliceVision.
• Updates CUDA compute capability defaults for CUDA 13 and removes use of a deprecated CUDA runtime field at compile time.
• Adjusts OpenMP option handling for AppleClang and adds an assimp/draco patch step for GCC 13.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 5 comments.

File	Description
src/cmake/Dependencies.cmake	Adds CUDA arch/CC configuration knobs, introduces GCC13 workaround patch for assimp’s bundled draco, tweaks superbuild configure flags.
src/aliceVision/gpu/gpu.cpp	Guards printing clockRate because it’s removed in CUDA 13.
src/CMakeLists.txt	Extends the default CUDA CC list for CUDA 13 (adds new architectures).
CMakeLists.txt	Changes how OpenMP default/disable behavior is expressed for AppleClang.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

-DALICEVISION_CUDA_CC_LIST=${AV_ALICEVISION_CUDA_CC_LIST} is always passed and has two concrete problems:

1. When AV_ALICEVISION_CUDA_CC_LIST is empty, this sets the cache entry to empty and can break later logic in src/CMakeLists.txt that assumes at least one CC (e.g., choosing the last entry for JIT flags).
2. If the user provides multiple CCs (e.g. 90;120), the semicolon will be treated as a list separator in CMake argument lists, splitting the -D... into multiple arguments and breaking the configure command.
   Only add this -D argument when non-empty, and pass it as a single escaped :STRING value so semicolons survive intact.

[Copilot]

The PATCH_COMMAND sed -i ... is not portable across platforms (BSD sed on macOS requires -i '', and some environments may not have GNU sed). Consider using a CMake-script patch step (e.g., PATCH_COMMAND ${CMAKE_COMMAND} -P ...) or a patch file applied via cmake -E/patch so the superbuild works consistently on non-Linux systems that may still use GCC 13.

[Copilot]

With CUDA 13 you add CC values 100 and 120, but the build later uses list(SORT ALICEVISION_CUDA_CC_LIST) (string sort) and then picks the last element as the highest CC for JIT compilation. String sorting will mis-order 3-digit CCs (e.g., 90 sorts after 120), causing the wrong "highest" CC to be selected. Sort using natural/numeric ordering (e.g., list(SORT ... COMPARE NATURAL)), or compute the numeric max explicitly.