Missing -fPIC in nvcc host compiler.

[bartoszek]

Enconture small problem while building dense_stereo on linux (arch_linux,gcc_7.1,cmake_3.8.1,cuda_8.0,cuda_host_compiler=gcc5_5.4.0)
linker complains about missing -fPIC flag in libmvs.a

/usr/bin/ld: ../mvs/libmvs_cuda.a(mvs_cuda_generated_patch_match_cuda.cu.o): relocation R_X86_64_32S against `.rodata' can not be used when making a shared object; recompile with -fPIC
/usr/bin/ld: ../mvs/libmvs_cuda.a(mvs_cuda_generated_gpu_mat_prng.cu.o): relocation R_X86_64_32 against `.bss' can not be used when making a shared object; recompile with -fPIC
/usr/bin/ld: ../mvs/libmvs_cuda.a(mvs_cuda_generated_gpu_mat_ref_image.cu.o): relocation R_X86_64_32S against `.rodata' can not be used when making a shared object; recompile with -fPIC
/usr/bin/ld: final link failed: Nonrepresentable section on output
collect2: error: ld returned 1 exit status
make[2]: *** [src/exe/CMakeFiles/dense_stereo.dir/build.make:163: src/exe/dense_stereo] Błąd 1
make[1]: *** [CMakeFiles/Makefile2:441: src/exe/CMakeFiles/dense_stereo.dir/all] Błąd 2
make[1]: *** Oczekiwanie na niezakończone zadania....

Adding --compiler-options -fPIC to CUDA_NVCC_FLAGS fixes the issue for me.

[ahojnnes]

Thanks for the feedback. Did you change anything in CMake or why are you building shared library objects?

[bartoszek]

I'm using default arch_linux build environment. The problem may relate to CMake=3.8 FindCUDA appending -fPIC to CFLAGS. ( Nope.. without cuda installed `-fPIC' still get added across the board ).
What I found in cmake build directory.

CMakeFiles/mvs_cuda.dir/flags.make:CXX_FLAGS = -march=x86-64 -mtune=generic -O2 -pipe -fstack-protector-strong -g -fvar-tracking-assignments -g -fvar-tracking-assignments -std=c++11 -Wno-maybe-uninitialized -fopenmp -O3 -DNDEBUG -fPIC   -Wall -pedantic` 

CMakeFiles/mvs.dir/flags.make:CXX_FLAGS = -march=x86-64 -mtune=generic -O2 -pipe -fstack-protector-strong -g -fvar-tracking-assignments -g -fvar-tracking-assignments -std=c++11 -Wno-maybe-uninitialized -fopenmp -O3 -DNDEBUG -fPIC   -Wall -pedantic -fPIC -std=gnu++11

Another possible cnadidate for forcing fPIC could be QT5:
/usr/lib/cmake/Qt5Core/Qt5CoreConfigExtras.cmake

 57 # Targets using Qt need to use the POSITION_INDEPENDENT_CODE property. The
 58 # Qt5_POSITION_INDEPENDENT_CODE variable is used in the # qt5_use_module
 59 # macro to add it.
 60 set(Qt5_POSITION_INDEPENDENT_CODE True)
 61
 62 # On x86 and x86-64 systems with ELF binaries (especially Linux), due to
 63 # a new optimization in GCC 5.x in combination with a recent version of
 64 # GNU binutils, compiling Qt applications with -fPIE is no longer
 65 # enough.
 66 # Applications now need to be compiled with the -fPIC option if the Qt option
 67 # "reduce relocations" is active. For backward compatibility only, Qt accepts
 68 # the use of -fPIE for GCC 4.x versions.
 69 if (CMAKE_VERSION VERSION_LESS 2.8.12
 70         AND (NOT CMAKE_CXX_COMPILER_ID STREQUAL "GNU"
 71         OR CMAKE_CXX_COMPILER_VERSION VERSION_LESS 5.0))
 72     set_property(TARGET Qt5::Core APPEND PROPERTY INTERFACE_POSITION_INDEPENDENT_CODE "ON")
 73 else()
 74     set_property(TARGET Qt5::Core APPEND PROPERTY INTERFACE_COMPILE_OPTIONS -fPIC)
 75 endif()
 76

As I can understand, qt is forcing fpic when building with gcc<=5.0, which is current host compiler for cuda on arch.

[ahojnnes]

Can you check whether adding list(APPEND CUDA_NVCC_FLAGS "--compiler-options -fPIC") after line 175 in the CMakeLists.txt file resolves the problem? Thanks!

[bartoszek]

Works here.

[ahojnnes]

Thanks, fixed in the dev branch now.