Fat binary with CUDA backend

[vitduck]

Hi,

To generate fat binary for multiple sm_*, I have tried two approaches as follow without success:

clang++ 
    -fsycl
    -fsycl-targets=nvptx64-nvidia-cuda-sycldevice
    -Xsycl-target-backend "--cuda-gpu-arch=sm_35" 
    -Xsycl-target-backend "--cuda-gpu-arch=sm_70" 
    vector_add.cpp -o vector_add.x 

and

clang++ 
    -fsycl
    -fsycl-targets=nvptx64-nvidia-cuda-sycldevice
    -Xsycl-target-backend "--cuda-gpu-arch=sm_35 --cuda-gpu-arch=sm_70" 
    vector_add.cpp -o vector_add.x 

I thought the latter were correct. However, the errer was

clang-13: error: cannot find libdevice for unknown. 
Provide path to different CUDA installation via --cuda-path, or pass -nocudalib to build without linking with libdevice.

Without '-Xsycl-target-backend', it works on V100, but not K40.
Please kindly advise the correct way to generate fat binary.
Thanks!

[alexbatashev]

This piece of code suggests, that at least CUDA plugin is not ready for what you're looking for. I guess, driver support is also limited in this scenario.
Also tagging @steffenlarsen.

[steffenlarsen]

There is an open PR for enabling this (#2087). Sadly I don't think any of the people behind it have the bandwidth to maintain it at this time.

[vitduck]

@alexbatashev , @steffenlarsen
Thanks for the clarification.
I will adjust the Makefile to accommodate separate SM.