[BUG] Fixed ARM Builds

[johnnynunez]

FIXES: #7130
ARM64 Linux builds are not available due to a runtime dynamic linker error: "cannot allocate memory in static TLS block"

Context:

Error migrate from 18.04 to 20.04 with glibc bugged.

Fixed:

Arm builds come back to build again without LTS Block problem.

Includes CUDA ARM Working again with >=12.1 cuda. Tested on Jetson and GH200

# On some aarch64 systems, libgfortran.a is not compiled with -fPIC,
# which prevents it from being used in a shared library.
# We link the shared version (-lgfortran) instead. This also handles libgcc.

[update-docs]

Thanks for submitting this pull request! The maintainers of this repository would appreciate if you could update the CHANGELOG.md based on your changes.

[johnnynunez]

@ssheorey merge!

[ssheorey]

Hi @johnnynunez you have 3 2 PRs open for fixing ARM builds. Which one did you want to merge? Please close the others.
I've triggered CI - please look for any build / test errors.

[johnnynunez]

Hi @johnnynunez you have 3 2 PRs open for fixing ARM builds. Which one did you want to merge? Please close the others. I've triggered CI - please look for any build / test errors.

Hi @johnnynunez you have 3 2 PRs open for fixing ARM builds. Which one did you want to merge? Please close the others. I've triggered CI - please look for any build / test errors.

the problem is gfortran. The another one opened is with all fixes that i did to build for ubuntu 22.04 and 24.04.

so gfortran is not build with shared files that is why is TLS Memory

[ssheorey]

Thanks @johnnynunez . Splitting into this smaller PR makes review much easier.

Other than the gfortran issue, there's one CI error in Ubuntu wheel. Perhaps something in cutlass was moved?

 #30 5863.6 In file included from /root/Open3D/cpp/open3d/ml/pytorch/continuous_conv/ContinuousConvBackpropFilterOpKernel.cu:11:
#30 5863.6 /root/Open3D/cpp/open3d/ml/impl/continuous_conv/ContinuousConvBackpropFilter.cuh:11:10: fatal error: cutlass/gemm/gemm.h: No such file or directory
#30 5863.6    11 | #include <cutlass/gemm/gemm.h>
#30 5863.6       |          ^~~~~~~~~~~~~~~~~~~~~
#30 5863.6 compilation terminated.
#30 5863.6 make[3]: *** [cpp/open3d/ml/pytorch/CMakeFiles/open3d_torch_ops.dir/build.make:609: cpp/open3d/ml/pytorch/CMakeFiles/open3d_torch_ops.dir/continuous_conv/ContinuousConvBackpropFilterOpKernel.cu.o] Error 1

[johnnynunez]

Thanks @johnnynunez . Splitting into this smaller PR makes review much easier.

Other than the gfortran issue, there's one CI error in Ubuntu wheel. Perhaps something in cutlass was moved?

 #30 5863.6 In file included from /root/Open3D/cpp/open3d/ml/pytorch/continuous_conv/ContinuousConvBackpropFilterOpKernel.cu:11:
#30 5863.6 /root/Open3D/cpp/open3d/ml/impl/continuous_conv/ContinuousConvBackpropFilter.cuh:11:10: fatal error: cutlass/gemm/gemm.h: No such file or directory
#30 5863.6    11 | #include <cutlass/gemm/gemm.h>
#30 5863.6       |          ^~~~~~~~~~~~~~~~~~~~~
#30 5863.6 compilation terminated.
#30 5863.6 make[3]: *** [cpp/open3d/ml/pytorch/CMakeFiles/open3d_torch_ops.dir/build.make:609: cpp/open3d/ml/pytorch/CMakeFiles/open3d_torch_ops.dir/continuous_conv/ContinuousConvBackpropFilterOpKernel.cu.o] Error 1

I will check. Yeah, maybe they moved internally in cutlasss. I did’t check ML modules. So i will check it asap.
It is important to migrate new cutlass and openblas for new arm devices