Fix nvrtc resolution when CUDA_HOME env is set

[ZzEeKkAa]

libnvrtc.so is located in the lib64 folder not in the separate folder. I guess it is a typo based on the nvvm location. Not sure about Windows path, but was following same logic as nvvm does, but without subfolder.

UPD: probably won't be needed after #308

[copy-pr-bot]

Auto-sync is disabled for ready for review pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[brandon-b-miller]

Thanks @ZzEeKkAa , is there a container or something similar I can use to reproduce the failure state?

[ZzEeKkAa]

Thanks @ZzEeKkAa , is there a container or something similar I can use to reproduce the failure state?

I was using just system install of ctk at /usr/local/cuda-X.Y. I would expect that any cuda container, for example https://hub.docker.com/r/rapidsai/devcontainers will work.

I was following this instructions to setup system ctk:

export CUDA_HOME=/usr/local/cuda-X.Y
export LIBRARY_PATH=$CUDA_HOME/lib64:$LIBRARY_PATH

[gmarkall]

This looks like it was incorrect previously, and that this makes it correct (I have not tested it, beyond inspecting where the NVRTC lib appears to be placed in a toolkit installation). I don't think we do any testing with CUDA_HOME, so I can imagine this could have been non-working since it was last modified.

[gmarkall]

On main, when I run:

CUDA_HOME=/usr/local/cuda-12.8 NUMBA_CUDA_USE_NVIDIA_BINDING=0 gdb --args python -c "from numba import cuda; cuda.cudadrv.libs.test(); input()"

I get

Finding driver from candidates:
	libcuda.so
	libcuda.so.1
	/usr/lib/libcuda.so
	/usr/lib/libcuda.so.1
	/usr/lib64/libcuda.so
	/usr/lib64/libcuda.so.1
Using loader <class 'ctypes.CDLL'>
	Trying to load driver...	ok
		Loaded from libcuda.so
	Mapped libcuda.so paths:
		/usr/lib/x86_64-linux-gnu/libcuda.so.570.133.20
Finding nvvm from CUDA_HOME
	Located at /usr/local/cuda-12.8/nvvm/lib64/libnvvm.so.4.0.0
	Trying to open library...	ok
Finding nvrtc from CUDA_HOME
	Located at libnvrtc.so
	Trying to open library...	ok
Finding cudadevrt from CUDA_HOME
	Located at /usr/local/cuda-12.8/lib64/libcudadevrt.a
	Checking library...	ok
Finding libdevice from CUDA_HOME
	Located at /usr/local/cuda-12.8/nvvm/libdevice/libdevice.10.bc
	Checking library...	ok
Include directory configuration variable:
	CUDA_INCLUDE_PATH=/usr/local/cuda/include
Finding include directory from CUDA_INCLUDE_PATH Config Entry
	Located at /usr/local/cuda/include
	Checking include directory...	ok

The nvrtc path was not located correctly - gdb info shared shows it found another nvrtc.so on the linker path:

0x00007ffef6352e10  0x00007ffef9654342  Yes (*)     /usr/local/cuda-12.8/nvvm/lib64/libnvvm.so.4.0.0
0x00007ffef2643e80  0x00007ffef4c9fc92  Yes (*)     /usr/local/cuda-12.6/targets/x86_64-linux/lib/libnvrtc.so

(nvvm.so also shown for comparison).

With this branch, the correct nvrtc library is loaded:

0x00007ffef6352e10  0x00007ffef9654342  Yes (*)     /usr/local/cuda-12.8/nvvm/lib64/libnvvm.so.4.0.0
0x00007ffeefe1f170  0x00007ffef46b2892  Yes (*)     /usr/local/cuda-12.8/lib64/libnvrtc.so.12.8.93

and the output is also as expected:

Finding nvrtc from CUDA_HOME
	Located at /usr/local/cuda-12.8/lib64/libnvrtc.so.12.8.93
	Trying to open library...	ok

[gmarkall]

/ok to test