Detect gpu architecture in build.rs and pass --gpu-architecture flag to nvcc

[coreylowman]

Unsure how to detect as of now, so part of this issue is to figure that out!

[Narsil]

nvidia-smi --query-gpu=compute_cap --format=csv

Or

cudaDeviceProp deviceProp;
cudaGetDeviceProperties(&deviceProp, dev);
std::printf("%d.%d\n", deviceProp.major, deviceProp.minor)

Might be better though (nvidia-smi is not necessarily installed)
https://stackoverflow.com/questions/48283009/nvcc-get-device-compute-capability-in-runtime

[JoeOsborn]

Two questions on this:

1. Is -arch=native what is being asked for here? (Per https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html#options-for-steering-gpu-code-generation )
2. Would -arch=all or -arch=all-major be a better choice for release builds?

Otherwise you might need to get the device name(s) and check them against this list or something: https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html#gpu-feature-list

[Narsil]

Would -arch=all or -arch=all-major be a better choice for release builds?

I think not, one asset of being a compiled framework, is that we can make it highly optimized for the highest possible compute capability making small and efficient binaries for a specific platform it's supposed to work on.

Making this overridable for users that want to create shared binaries would be good. But I don't think it should be the default.

[coreylowman]

Oh -arch=native seems exactly what we want! Easy change too, thanks for sharing that