Error when compiling kernel (using the example python script) about deepdrr HOT 3 CLOSED

We have overhauled the Projector to use cupy, since Textures are officially deprecated with CUDA 12 and their replacement is not supported by PyCUDA. We will push this release after further testing, but if you would like to try it out, you can check out the dev branch and install from source.

from deepdrr.

The error stems from the "texture" type being deprecated since CUDA 5.0 1. Support was entirely dropped in version 12.0.

deepdrr/projector/cubic/cubicTex3D.cu(54): error: identifier "texture" is undefined __attribute__((device)) float linearTex3D(texture<T, 3, mode> tex, float3 coord)

This issue is quite nasty to refactor, as the API significantly changed. Grepping for "texture<", I found about 360 occurrences in the whole DeepDRR project. Further, I learned that the project uses third party code from 2010 for performing cubic interpolation (http://www.dannyruijters.nl/cubicinterpolation/) which relies on texture objects.

I am not very familiar with CUDA programming, but I see the following options to resolve:

• Refactor the existing code to use the new cudaTextureObject_t API, maybe under the help of a LMM.
• Replace the CUDA kernels by an abstraction layer like cupy, which is basically equal to a rewrite.
• Ask users to downgrade CUDA 11.x. However, this is not a sustainable solution and might not be possible in many setups (e.g. for students who don't have admin rights on the university machines).

from deepdrr.

This of course is unfortunate. We are reviewing this, and will likely be implementing a change. @benjamindkilleen and @liamjwang may have additional thoughts/updates on this now or soon.

from deepdrr.