with cuda 12.2.0-1, when running /opt/cuda/integration/nsight-compute/ncu-ui, it will says

$ /opt/cuda/integration/nsight-compute/ncu-ui 
ERROR : nsight-compute directory is not found under /opt/cuda/integration/nsight-compute/../ or /opt/nvidia. Nsight Compute is not installed on your system.

the nsight is in one split package cuda-tools and its "Package Contents" has "opt/cuda/nsight_compute/"

Feel free to adopt the package

The PKGBUILD part is actually correct and can still work. Modifications need to do:

• Change the pkgver

• Login to NVIDIA website, download the .run file manually from https://developer.nvidia.com/nsight-graphics , change the first source into

file:///${pkgname}_${pkgver}.run

• Delete the part under Fix install path /usr to ${pkgdir}/usr for registering of layers

  # sed -i "s,\([= ]\)/usr/,\1${pkgdir}/usr/,g" ${srcdir}/${pkgname}/pkg/host/linux-desktop-nomad-x64/VK_LAYER_NV_nomad.sh
  # sed -i "s,\([= ]\)/usr/,\1${pkgdir}/usr/,g" ${srcdir}/${pkgname}/pkg/host/linux-desktop-nomad-x64/VK_LAYER_NV_GPU_Trace.sh

and the same content under package()

  # rm ${pkgdir}/opt/${pkgname}/host/linux-desktop-nomad-x64/VK_LAYER_NV_nomad.sh
  # rm ${pkgdir}/opt/${pkgname}/host/linux-desktop-nomad-x64/VK_LAYER_NV_GPU_Trace.sh

Nsight Compute is shipped together with CUDA, but it can also be downloaded as a standalone; they also sometimes release updates to it in between CUDA releases, so one might want to go and grab the latest update directly from the website in those cases. And that standalone download has been requiring a login for at least a year or two IIRC.

Sounds like the DLAGENTS thing could be an interesting work around.

I might be able to get registered users to use the package without having to download the installer like in optix.

I do some DLAGENTS hacking in my gog-* packages and I could add a shell or python script to ask for credentials once, store them locally and use that to download. This would be a last resort. That or like what optix does.

The thing is that web requests aren't my expertise so I can figure it out but it will take me some time to find the proper requests to make to get the tokens needed. Also, it gets much more complicated with 2FA involved.

Ideally, NVIDIA could reverse their change. It could be an oversight. I emailed them to let them know. You can always respectfully let them know your thoughts on their contact page [1].

[1] https://developer.nvidia.com/contact

I think nsight compute comes with the cuda package. Interestingly, you don't need to be registered to download that.

This is unfortunate, but not surprising as some other tools (Nsight Compute maybe?) had already been requiring to login before downloading for some time.

I am guessing something similar to the optix AUR package will be needed, where the user downloads the script themselves and place it at a specific location for the package to use.

Maybe contact Nvidia to reverse that?

They probably won't (unfortunately) but they should understand that blocking direct access is stupid

Looks like nvidia started requiring users login via the NVIDIA Registered Developer Program to download their installer. I don't know if there's anything I can do about that.