[CMAKE] CUDA virtual architectures support ? about opencv HOT 5 CLOSED

Currently OpenCV does not use CUDA_ARCHITECTURES (I am almost 100% sure this requires CUDA to be a first class language). Instead it can do one of four things, build for

1. all architechtures supported by the installed version of CUDA toolkit (almost equivelent to all) if CUDA_ARCH_BIN and CUDA_ARCH_PTX are not specified, or
2. architectures specified by combinations of CUDA_ARCH_BIN and CUDA_ARCH_PTX, or
3. a specific generation, e.g. CUDA_GENERATION=Lovelace, or
4. the architechtures of the hosts GPUs if CUDA_GENERATION=Auto (equivelent to native).

If CUDA is a added as a first class language (#23021) these options could be supported however they would not enable OpenCV to determine which real and virtual architechtures it was being built for which is required for both cuda::TargetArchs at runtime and the information shown by CMake at configuration time e.g.

-- NVIDIA GPU arch: 50 52 60 61 70 75 80 86 89 90
-- NVIDIA PTX archs: 90

from opencv.

Ok, it's more clear.

But is it documented somewhere ? The tooltips in CMakeGUI above CUDA_GENERATION are not very informative.

from opencv.

But is it documented somewhere ?

I don't think so apart from a mention here.

from opencv.

Should I suggest a doc improvement (for instance a URL in the CMake tooltip pointing to an explanation in a dedicated page), leaving this thread open as a reminder, or do I close it ?

from opencv.

Should I suggest a doc improvement (for instance a URL in the CMake tooltip pointing to an explanation in a dedicated page),

You could add the information to the CUDA Module Introduction page however that page is really out of date and all the information could do with an update.

from opencv.