System setup

x86_64:

• Lenovo Legion 5i
• GPU: NVidia RTX 3070
• Driver Version: 580.126.09
• CUDA Version: 13.0
• Ubuntu Jammy 22.04

arm64:

• Jetson Orin Nano
• Driver Version: 540.5.0
• CUDA Version: 12.6
• Ubuntu Jammy 22.04
• Jetpack (6.2.2)Version: 6.2.2+b24

Description

• Building cuVSLAM natively on Jetson required several undocumented steps and workarounds.
• The hardcoded -arch=all bypassed CMake's architecture management, and there was no way to target a specific GPU from the build script.
• The speedup test filter bug caused false test failures.

To Reproduce

• build the libcuvslam for target architecture of the host or for multitude of architectures
  ./build_release.sh --cuda_arch=87 --modules_test --build_lib - to target specific arch
  ./build_release.sh --modules_test --build_lib - to target multiple arch
• build the module tests
  ./build_release.sh --modules_test

Expected behavior

• A clean build for target architecture when required
• A clean multi-arch build when no specific target architecture is passed
• And all the 14 module tests passed

Observed behavior

• current build doesn't accept a param for passing the target cuda arch and has a CUDA_ARCH=all hardcoded for building libcuVSLAM for all archs.
• module tests produce error as the case-sensitive filter was missing the image_pyramid_test speedup variable

Additional information
PR #38 fixes this