CUDA on ARM

[totaam]

CUDA on ARM is a thing: CUDA is now supported on Arm servers starting with CUDA 11

Remains to be seen if we can enable nvfbc, nvenc and nvjpeg builds on arm.

[totaam]

The build scripts have been updated with a per-arch CUDA directory: Xpra-org/repo-build-scripts@e12d874 so the commit above tries to enable building nvjpeg and nvenc.
nvfbc is left disabled for now because it looks like this isn't supported on arm64: the library is usually installed as part of the driver and I can't seem to find it in the arm64 installation.

[totaam]

Building on aarch64 with the updated build scripts looks like this:

* nvenc                : Y
* nvfbc                : N
* nvjpeg               : Y
(..)
found CUDA compiler: /opt/cuda-arm64/bin/nvcc version 11.5.119
rebuilding XRGB_to_NV12: no file
CUDA compiling XRGB_to_NV12     (no file)
 '/opt/cuda-arm64/bin/nvcc' \
    '-fatbin' \
    '-c' 'fs/share/xpra/cuda/XRGB_to_NV12.cu' \
    '-o' 'fs/share/xpra/cuda/XRGB_to_NV12.fatbin' \
    '-std=c++03' \
    '-gencode=...  ..arch=compute_86,code=sm_86'

[totaam]

To be able to compile pycuda, I needed to add DefaultTasksMax=unlimited to /etc/systemd/system.conf.

[totaam]

It took many hours to build the packages, especially pycuda and xpra.
But we now have nvenc and nvjpeg support for arm64 in the Fedora repositories.
(totally untested as I don't have any arm64 hardware with PCIE)