linker error when building with libPhysXGpu_64 binary.

[noblegmr]

Hi everyone!

I am currently working on a project where I am creating SDFs' for meshes (more information about that here). For this task I want to use the GPU capability for creating SDFs that Physx has to offer.
To do this I need to access a CudaManager object. which relies on the PxGpu.h and PxPhysicsGpu.h files. These file is a part of the libPhysXGpu_64.so static file. (attaching code snippet to setup_SDFDesc.txt).

Context on environment:
I am running this inside a docker container environment, with CUDA version 12.2 installed and Physx 5.4.0 installed. I followed the following steps inside docker container to install and setup the Physx libraries (attached snippet in physx_setup.txt file).
And for the CMAKE list for my project. I included the addresses of all the static binaries (attached in cmake_setup.txt).

I can see all the respective static libraries (including the libPhysXGpu_64) inside the /PhysX/physx/bin/linux.clang/checked/ folder. but I still ended up getting the following linker error (attached as linker_error1.txt).

I checked out a related issue here: and tried out the solution in moving the .so file into the correct directory, but that still did not work instead I got a different error where all the functions inside the binary were undefined (attached as linker_error2.txt).
Please do let me know if anyone has any ideas or suggestions! thank you!

setup_SDFDesc.txt
physx_setup.txt.
linker_error1.txt
linker_error2.txt
physx_cmake.txt

[Spatchler]

Have you been able to fix this?

[noblegmr]

Hi this is still an issue, I've just been using the CPU SDF builder for the time being. currently Ironing out some other issues with SDF usage.

[preist-nvidia]

@noblegmr - this is a mistake on our side, we accidentally built the GPU shared library on Ubuntu 22 resulting in a glibc incompatiblity with Ubuntu 20 builds. This is fixed in 5.4.1, please update to this release. Apologies for the trouble!

[noblegmr]

Hi, thank you for your reply. I will try this out and let you know how it goes! thank you

[rishavgmr]

@preist-nvidia We have tried with the newer version of PhysX (5.4.2) but the issue still persists with the usage of libPhysXGpu_64.so. When trying to link against it, we keep getting this error:

/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuArray3DCreate_v2'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuStreamQuery'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuDeviceGet'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuTexObjectCreate'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuStreamDestroy_v2'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuStreamSynchronize'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuStreamWaitEvent'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuCtxGetStreamPriorityRange'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuEventSynchronize'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuMemsetD32_v2'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuCtxDestroy_v2'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuMemcpyPeerAsync'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuEventDestroy_v2'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuEventCreate'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuMemGetInfo_v2'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuLaunchKernel'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuStreamCreate'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuInit'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuEventRecord'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuDeviceGetAttribute'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuCtxPopCurrent_v2'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuModuleUnload'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuCtxGetDevice'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuDeviceGetName'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuTexObjectDestroy'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuMemFreeHost'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuMemcpyDtoHAsync_v2'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuMemcpyHtoD_v2'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuCtxPushCurrent_v2'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuMemFree_v2'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuEventQuery'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuMemcpyDtoH_v2'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuCtxCreate_v2'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuStreamCreateWithPriority'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuMemcpyHtoDAsync_v2'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuMemAlloc_v2'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuMemHostGetDevicePointer_v2'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuModuleGetFunction'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuMemcpyDtoD_v2'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuModuleLoadDataEx'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuDeviceTotalMem_v2'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuMemsetD8_v2'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuArrayDestroy'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuMemcpyDtoDAsync_v2'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuMemsetD32Async'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuMemcpy3DAsync_v2'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuMemsetD16_v2'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuMemHostAlloc'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuDriverGetVersion'
/usr/bin/ld: /opt/PhysX/physx/install/linux/PhysX/bin/linux.clang/checked/libPhysXGpu_64.so: undefined reference to `cuMemsetD8Async'

I have verified that CUDA is installed on the system by running nvcc --version and the CMakeLists.txt for the project is the same as the one shown in the top level question on this thread physx_cmake.txt. Is there anything we need to do build PhysX a certain way such that we can use the GPU functions?

[rishavgmr]

The linker errors go away after adding following to the target_link_libraries:

${CUDA_LIBRARIES}
cuda

In total, the whole CMakeFile should be something like this: physx_cmake_correct.txt

We added the find_package for CUDA and the associated CMAKE variables for the includes and libraries. Need to make special node of adding cuda to the target_link_libraries in addition to ${CUDA_LIBRARIES}.

I will test now to see if it is actually using the GPU by checking for a speed up in SDF calculation time.

[rishavgmr]

This seems to have fixed the issue. It seems to be actually using the GPU to calculate the SDF since we can see a 5x speed up compared to CPU.

[preist-nvidia]

Great!