CERN/TIGRE

Tigre could not compile under CUDA 12.0

yliu88au opened this issue · 5 comments

Expected Behavior

Actual Behavior

Two issues:
(1) Need remove support for earlier GPU type "compute_37" to run setup.py
(2) Error compiling some *.cu code, for example
../Common/CUDA/ray_interpolated_projection_parallel.cu(70): error: texture is not a template

Looks like CUDA 12.0 does not support texture reference any more. See
https://forums.developer.nvidia.com/t/cuda-12-0-still-support-for-texture-reference-support-for-pascal-architecture-warp-synchronous-programming/237284

Code to reproduce the problem (If applicable)

python setup.py install


Specifications

python version: 3.9.13

  • OS: Ubuntu 20,04
  • CUDA version: 12.0

Indeed, I am aware of this, but thanks for opening an issue.

Both are solvable issues.

1- setup.py (and the equivalent Compile.m) both already do modify the compilation flags given the CUDA selected. These functions need to be simply updated to account for CUDA 12.
2-Most TIGRE already uses texture objects, but some of the older code (parallel beam) still uses texture references. Modifying this should hopefully not be a big deal.

I will try to find time to do these, but can't promise to fix them anytime "soon".

I have now updated master to have CUDA 12 support, but I don't have the capability to test it.

I someone does, please let me know.

tested and working.

Fantastic!

I think that warning is OK, as those files are just for compiling, so they don't need to be a package.
Thanks for double testing, I appreciate it!