Segmentation fault when building with VTKm

ParaView Support
1

Hello!

I’m trying to build ParaView with VTKm enabled but all binaries (paraview, pvpython, and servers) failed with SIGSEGV (segmentation fault) when I start the program, in the initialization without any message, shell or window.
Without VTKm but with CUDA/OptiX the binaries run fine. Other CUDA apps that I build also run fine so I think my CUDA install is alright.

I tried to change the VTKm_CUDA_Architecture from “native” to “pascal” (using a 1060 card here) but both give the SIGSEGV, tried enabling CUDA_SEPARABLE_COMPILATION with no success, and tried to enable Debug but it has not improved the backtrace.

I found some mention of a similar error in a CMake merge request (https://gitlab.kitware.com/cmake/cmake/merge_requests/457) but don’t know if it’s related, maybe I need to manually set the CMAKE_CUDA_FLAGS like “-gencode arch=compute_61,code=sm_61” (I use this on my Ph.D. project with success).

Using CMake 3.14.6, CUDA 10.0.130 (with GCC 7.4.0), GCC 9.2.0 for c++ code and NVidia driver 440.31. If more information is needed please let me know.

Can anyone point me in the direction to fix this?

The debug backtrace from cuda-gdb with CMAKE_BUILD_TYPE=Debug is:

0x00007ffff359139a in __cudaRegisterLinkedBinary(__fatBinC_Wrapper_t const*, void (*)(void**), void*) ()
   from /root/src/paraview_build/bin/../lib64/libvtkPVServerManagerApplication-pv5.7.so.1
(cuda-gdb) bt
#0  0x00007ffff359139a in __cudaRegisterLinkedBinary(__fatBinC_Wrapper_t const*, void (*)(void**), void*) ()
   from /root/src/paraview_build/bin/../lib64/libvtkPVServerManagerApplication-pv5.7.so.1
#1  0x00007ffff3590380 in __cudaRegisterLinkedBinary_48_tmpxft_0000622a_00000000_6_ClipWithField_cpp1_ii_7724b337 ()
   from /root/src/paraview_build/bin/../lib64/libvtkPVServerManagerApplication-pv5.7.so.1
#2  0x00007ffff421670d in __sti____cudaRegisterAll ()
   from /root/src/paraview_build/bin/../lib64/libvtkPVServerManagerApplication-pv5.7.so.1
#3  0x00007ffff7fe318a in _dl_rtld_di_serinfo () from /lib64/ld-linux-x86-64.so.2
#4  0x00007ffff7fe3286 in _dl_rtld_di_serinfo () from /lib64/ld-linux-x86-64.so.2
#5  0x00007ffff7fd50ca in ?? () from /lib64/ld-linux-x86-64.so.2
#6  0x0000000000000001 in ?? ()
#7  0x00007fffffffe28b in ?? ()
#8  0x0000000000000000 in ?? ()

Thanks for your time!

2

@robert.maynard @allison.vacanti

3

Does this occur if you have VTK-m enabled, and OptiX disabled?

Edit:
Follow-up. You are building ParaView master?

4

I’ve tried a clean build with only VTK-m enabled and it worked fine, even with PARAVIEW_USE_CUDA=ON. Then just enabling ray tracing and VisRTX worked fine with not SIGSEGV.

I’m using master @697b01a200.

It may be related to some flag I enabled on the previous build, or the fact now I didn’t “make clean” between builds, really don’t know.
Will try to reproduce the problem here.

5

This sounds like the classic issue where two dynamic libraries that are using static CUDA and therefore get multiple duplicate symbols.

I think the quickest solution is to add -Xlinker muldefs to you CMAKE_CUDA_FLAGS and see if that builds properly