Strange link error on ubuntu 22.04 about owl HOT 6 CLOSED

Looks like cudaGetDeviceProperties_v2 it actually missing in cuda toolkit

nm /usr/local/cuda-11.8/lib64/libcudart_static.a |grep cudaGetDeviceProperties
000000000003ec90 T cudaGetDeviceProperties

while it is ok in nvhpc

nm /data/pkestene/local/hpcsdk-23.3/Linux_x86_64/23.3/cuda/12.0/lib64/libcudart_static.a| grep cudaGetDeviceProperties
0000000000064010 T cudaGetDeviceProperties
000000000003e860 T cudaGetDeviceProperties_v2

it is also OK, when moving to cuda toolkit >=12.

from owl.

so just double-checking on this: this looks like you built with one cuda 11.8 install and then tried to run with another 11.8 install, and that was causing the issue? ie, both by themselves would have worked, just mixing them didn't?

(agreed that having to different distributions with the same version numbers if "funny", though :-) just trying to make sure that it's not related to owl.)

from owl.

Just to be clear:

• building with cuda 11.8 (from toolkit) fails, the error is at link
• building with cuda (from nvhpc, here 12.0) is fine, and I run using the same env
• building with cuda 12.0 (from toolkit) is fine, runs fine

The problem was that libcudart_static.a shipped with cuda 11.8 doesn't provide cudaGetDeviceProperties_v2.

At least, it's working now. My next step is playing with owlExaBrick.
Thanks for making this available.

from owl.

Huh; that is "slightly" concerning. Basically what you're saying is that CUDA 11.8 is broken :-/. Huh. Now we have three options: a) try and fix the code even for cuda 11.8; b) go into cmakefile, detect cuda version, and at least throw an error; or c) ignore, and hope that people will use the newer cuda 12, anyway...
Anyway - thanks for reporting this - OS/toolchain related stuff is always nasty, in particular when version dependent...

from owl.

I tried another machine under RedHat8, and cuda 11.8 (both toolkit and nvhpc), no problem there.

When I look for symbols cudaGetDeviceProperties/cudaGetDeviceProperties_v2 on that machine, I get the same results as on Ubuntu, that is:

> nm /ccc/products/cuda-11.8/system/toolkit/lib64/libcudart_static.a | grep cudaGetDeviceProper
000000000003ec90 T cudaGetDeviceProperties
> nm /ccc/products/cuda-12.0/system/toolkit/lib64/libcudart_static.a | grep cudaGetDeviceProper
0000000000064010 T cudaGetDeviceProperties
000000000003e860 T cudaGetDeviceProperties_v2

cuda-11.8 only contain one of them, cuda-12.0 contains them both.

But on RedHat, owl samples apps link fine with cuda-11.8, even though cudaGetDeviceProperties_v2 is not present.

from owl.

I finally found the problem on my ubuntu machine; eventhough both cuda toolkit 11.8 and 12.0 where installed, in complete separated directories, when installing newer toolkit, by default the ubuntu package creates a sym link /usr/local/cuda -> /etc/alternatives/cuda -> /etc/alternatives/cuda-12.0
so /usr/local/cuda always points to the latest cuda toolkit installed.

so what happened is that, I was compiling with nvcc 11.8, the cuda headers where actually taken from 12.0. So there was a mismatch between the header version and the runtime library version.

So the question is why is /usr/local/cuda/include included in CUDA_INCLUDES ?

Finaly, I think the problem is there:
https://github.com/owl-project/owl/blob/master/owl/CMakeLists.txt#L165

the path /usr/local/cuda/include is unconditionnaly included.

But this path is really not needed if using alias library like CUDA::cudart_static.

A possible fix is to replace:

target_include_directories(owl
  PUBLIC
    /usr/local/cuda/include/
    ${PROJECT_SOURCE_DIR}
    ${CMAKE_CURRENT_LIST_DIR}/include
)

by

target_include_directories(owl
  PUBLIC
    ${PROJECT_SOURCE_DIR}
    ${CMAKE_CURRENT_LIST_DIR}/include
)

So that the path /usr/local/cuda/include/ is not added by default.

I think that definitely close the issue. I can provide a small if needed.

from owl.