Unable to build OpenCV with CUDA on RHEL 7 - c++

Question

I am trying to build OpenCV tag 4.2 with CUDA 11.1 in an RHEL 7 docker image and are running into issues.
First issue/warning is this:
THRUST_COMPILER_DEPRECATION(C++14, pass -std=c++14 to your compiler);
Which can be resolved by removing --std=c++11 from cmake/OpenCVDetectCUDA.cmake:265.
Alright, simple enough to handle that warning. Although it would have been better to change it using a cmake option/define. The compile errors for gpu_mat.cu are there with or without the above change.
Then main problem is the build errors that are happening for gpu_mat.cu. Here is my current CMake line. I have tried a bunch of different options and at this point I'm just throwing everything and kitchen sink at it.
cmake ../ \
-DCMAKE_BUILD_TYPE=Release \
-DCMAKE_INSTALL_PREFIX=/usr \
-DENABLE_CXX11=OFF \ # I have tried it set to ON \
-DCMAKE_CXX_FLAGS="-std=c++14" \ # and without this line.
-DOPENCV_EXTRA_MODULES_PATH=../../opencv_contrib/modules/ \
-DWITH_EIGEN=OFF \
-DBUILD_TESTS=OFF \
-DBUILD_PERF_TESTS=OFF \
-DOPENCV_ENABLE_NONFREE:BOOL=ON \
-DWITH_CUDA:BOOL=ON \
-DCMAKE_PREFIX_PATH=/usr/local/cuda-11.1 \
-DCUDA_ARCH_BIN="52 60 61 70 75 80 86" \
-DCUDA_SEPARABLE_COMPILATION=ON \
-DCUDA_VERBOSE_BUILD=ON \
-DCUDA_NVCC_FLAGS="--std=c++14" \ # even passing this didn't resolve THRUST warning
-DCUDA_FAST_MATH=ON \
-DCUDA_PROPAGATE_HOST_FLAGS=OFF \
-DCUDA_BUILD_CUBIN=ON \
-DCUDA_HOST_COMPILER=/opt/rh/devtoolset-8/root/usr/bin/c++
Here is the VERBOSE output for gpu_mat.cu build:
[ 26%] Building NVCC (Device) object modules/core/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_gpu_mat.cu.o
cd /opt/tmp/opencv/build/modules/core/CMakeFiles/cuda_compile_1.dir/src/cuda && /usr/bin/cmake -E make_directory /opt/tmp/opencv/build/modules/core/CMakeFiles/cuda_compile_1.dir/src/cuda/.
cd /opt/tmp/opencv/build/modules/core/CMakeFiles/cuda_compile_1.dir/src/cuda && /usr/bin/cmake -D verbose:BOOL=ON -D build_configuration:STRING=Release -D generated_file:STRING=/opt/tmp/opencv/build/modules/core/CMakeFiles/cuda_compile_1.dir/src/cuda/./cuda_compile_1_generated_gpu_mat.cu.o -D generated_cubin_file:STRING=/opt/tmp/opencv/build/modules/core/CMakeFiles/cuda_compile_1.dir/src/cuda/./cuda_compile_1_generated_gpu_mat.cu.o.cubin.txt -P /opt/tmp/opencv/build/modules/core/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_gpu_mat.cu.o.Release.cmake
-- Removing /opt/tmp/opencv/build/modules/core/CMakeFiles/cuda_compile_1.dir/src/cuda/./cuda_compile_1_generated_gpu_mat.cu.o
/usr/bin/cmake -E rm -f /opt/tmp/opencv/build/modules/core/CMakeFiles/cuda_compile_1.dir/src/cuda/./cuda_compile_1_generated_gpu_mat.cu.o
-- Generating dependency file: /opt/tmp/opencv/build/modules/core/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_gpu_mat.cu.o.NVCC-depend
/usr/local/cuda-11.1/bin/nvcc -M -D__CUDACC__ /opt/tmp/opencv/modules/core/src/cuda/gpu_mat.cu -o /opt/tmp/opencv/build/modules/core/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_gpu_mat.cu.o.NVCC-depend -ccbin /opt/rh/devtoolset-8/root/usr/bin/c++ -m64 -D__OPENCV_BUILD=1 -D_USE_MATH_DEFINES -D__STDC_CONSTANT_MACROS -D__STDC_LIMIT_MACROS -D__STDC_FORMAT_MACROS -DOPENCV_WITH_ITT=1 --std=c++14 -gencode arch=compute_52,code=sm_52 -gencode arch=compute_60,code=sm_60 -gencode arch=compute_61,code=sm_61 -gencode arch=compute_70,code=sm_70 -gencode arch=compute_75,code=sm_75 -gencode arch=compute_80,code=sm_80 -gencode arch=compute_86,code=sm_86 -D_FORCE_INLINES --use_fast_math -Xcompiler -DCVAPI_EXPORTS -Xcompiler -fPIC -DNVCC -I/usr/local/cuda-11.1/include -I/opt/tmp/opencv/build/3rdparty/ippicv/ippicv_lnx/icv/include -I/opt/tmp/opencv/build/3rdparty/ippicv/ippicv_lnx/iw/include -I/opt/tmp/opencv/build -I/opt/tmp/opencv/modules/core/include -I/opt/tmp/opencv/build/modules/core -I/opt/tmp/opencv_contrib/modules/cudev/include -I/opt/tmp/opencv/3rdparty/include/opencl/1.2 -I/opt/tmp/opencv/3rdparty/ittnotify/include
-- Generating temporary cmake readable file: /opt/tmp/opencv/build/modules/core/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_gpu_mat.cu.o.depend.tmp
/usr/bin/cmake -D input_file:FILEPATH=/opt/tmp/opencv/build/modules/core/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_gpu_mat.cu.o.NVCC-depend -D output_file:FILEPATH=/opt/tmp/opencv/build/modules/core/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_gpu_mat.cu.o.depend.tmp -D verbose=ON -P /usr/share/cmake-3.18/Modules/FindCUDA/make2cmake.cmake
-- Copy if different /opt/tmp/opencv/build/modules/core/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_gpu_mat.cu.o.depend.tmp to /opt/tmp/opencv/build/modules/core/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_gpu_mat.cu.o.depend
/usr/bin/cmake -E copy_if_different /opt/tmp/opencv/build/modules/core/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_gpu_mat.cu.o.depend.tmp /opt/tmp/opencv/build/modules/core/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_gpu_mat.cu.o.depend
-- Removing /opt/tmp/opencv/build/modules/core/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_gpu_mat.cu.o.depend.tmp and /opt/tmp/opencv/build/modules/core/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_gpu_mat.cu.o.NVCC-depend
/usr/bin/cmake -E rm -f /opt/tmp/opencv/build/modules/core/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_gpu_mat.cu.o.depend.tmp /opt/tmp/opencv/build/modules/core/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_gpu_mat.cu.o.NVCC-depend
-- Generating /opt/tmp/opencv/build/modules/core/CMakeFiles/cuda_compile_1.dir/src/cuda/./cuda_compile_1_generated_gpu_mat.cu.o
/usr/local/cuda-11.1/bin/nvcc /opt/tmp/opencv/modules/core/src/cuda/gpu_mat.cu -dc -o /opt/tmp/opencv/build/modules/core/CMakeFiles/cuda_compile_1.dir/src/cuda/./cuda_compile_1_generated_gpu_mat.cu.o -ccbin /opt/rh/devtoolset-8/root/usr/bin/c++ -m64 -D__OPENCV_BUILD=1 -D_USE_MATH_DEFINES -D__STDC_CONSTANT_MACROS -D__STDC_LIMIT_MACROS -D__STDC_FORMAT_MACROS -DOPENCV_WITH_ITT=1 --std=c++14 -gencode arch=compute_52,code=sm_52 -gencode arch=compute_60,code=sm_60 -gencode arch=compute_61,code=sm_61 -gencode arch=compute_70,code=sm_70 -gencode arch=compute_75,code=sm_75 -gencode arch=compute_80,code=sm_80 -gencode arch=compute_86,code=sm_86 -D_FORCE_INLINES --use_fast_math -Xcompiler -DCVAPI_EXPORTS -Xcompiler -fPIC -DNVCC -I/usr/local/cuda-11.1/include -I/opt/tmp/opencv/build/3rdparty/ippicv/ippicv_lnx/icv/include -I/opt/tmp/opencv/build/3rdparty/ippicv/ippicv_lnx/iw/include -I/opt/tmp/opencv/build -I/opt/tmp/opencv/modules/core/include -I/opt/tmp/opencv/build/modules/core -I/opt/tmp/opencv_contrib/modules/cudev/include -I/opt/tmp/opencv/3rdparty/include/opencl/1.2 -I/opt/tmp/opencv/3rdparty/ittnotify/include
And a snippet of the errors that immedately follow:
/opt/rh/devtoolset-8/root/usr/include/c++/8/bits/functional_hash.h: In member function 'std::size_t std::hash<float>::operator()(float) const':
/opt/rh/devtoolset-8/root/usr/include/c++/8/bits/functional_hash.h:232:22: error: expected ')' before numeric constant
return __val != 0.0f ? std::_Hash_impl::hash(__val) : 0;
~ ^~~
)
/opt/rh/devtoolset-8/root/usr/include/c++/8/bits/functional_hash.h:232:15: error: invalid operands of types 'float' and 'double(const char*) throw ()' {aka 'double(const char*)'} to binary 'operator!='
return __val != 0.0f ? std::_Hash_impl::hash(__val) : 0;
~~~~~~^~~~~~~~~~
/opt/rh/devtoolset-8/root/usr/include/c++/8/bits/functional_hash.h:232:64: error: expected ')' before ';' token
return __val != 0.0f ? std::_Hash_impl::hash(__val) : 0;
~ ^
)
/opt/rh/devtoolset-8/root/usr/include/c++/8/bits/functional_hash.h: In member function 'std::size_t std::hash<double>::operator()(double) const':
/opt/rh/devtoolset-8/root/usr/include/c++/8/bits/functional_hash.h:244:22: error: expected ')' before numeric constant
return __val != 0.0 ? std::_Hash_impl::hash(__val) : 0;
~ ^~
)
/opt/rh/devtoolset-8/root/usr/include/c++/8/bits/functional_hash.h:244:15: error: invalid operands of types 'double' and 'double(const char*) throw ()' {aka 'double(const char*)'} to binary 'operator!='
return __val != 0.0 ? std::_Hash_impl::hash(__val) : 0;
.................................
~~~
/opt/tmp/opencv/modules/core/src/cuda/gpu_mat.cu:382:4: note: to match this '('
if (value[0] == 0.0 && value[1] == 0.0 && value[2] == 0.0 && value[3] == 0.0)
^
/opt/tmp/opencv/modules/core/src/cuda/gpu_mat.cu:415:14: error: 'func_t' does not name a type; did you mean 'fpos_t'?
static const func_t funcs[7][4] =
^~~~~~
fpos_t
/opt/tmp/opencv/modules/core/src/cuda/gpu_mat.cu:426:3: error: 'funcs' was not declared in this scope
funcs[depth()][channels() - 1](*this, value, stream);
^ ~~~
-- Removing /opt/tmp/opencv/build/modules/core/CMakeFiles/cuda_compile_1.dir/src/cuda/./cuda_compile_1_generated_gpu_mat.cu.o
/usr/bin/cmake -E rm -f /opt/tmp/opencv/build/modules/core/CMakeFiles/cuda_compile_1.dir/src/cuda/./cuda_compile_1_generated_gpu_mat.cu.o
CMake Error at cuda_compile_1_generated_gpu_mat.cu.o.Release.cmake:276 (message):
Error generating file
/opt/tmp/opencv/build/modules/core/CMakeFiles/cuda_compile_1.dir/src/cuda/./cuda_compile_1_generated_gpu_mat.cu.o
make[2]: *** [modules/core/CMakeFiles/opencv_core.dir/build.make:84: modules/core/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_gpu_mat.cu.o] Error 1
make[2]: Leaving directory '/opt/tmp/opencv/build'
make[1]: *** [CMakeFiles/Makefile2:3852: modules/core/CMakeFiles/opencv_core.dir/all] Error 2
make[1]: Leaving directory '/opt/tmp/opencv/build'
make: *** [Makefile:182: all] Error 2
I am running the build process after enabling devtools
scl enable devtoolset-8 bash # also tried with devtoolset-7
I have searched for these errors, but I guess my google-fu is weak for this. I didn't have these errors/problems when I built OpenCV 4.5.1 w/ CUDA in an Ubuntu 18 Docker image.
So I'm totally missing something and have been working on this for a couple of days. I am just out of ideas of what to do next.
Thanks for any input.

Answer 1

I'm seeing some very similar build errors as you are, so you're not alone! I'm told it has something to do with a recent change to centos 7, which has broken nvcc. Hopefully, it gets fixed soon.
https://gitlab.com/nvidia/container-images/cuda/-/issues/109#note_503061879

Answer 2

There was a reported issue with RHEL/CentOS and has since been resolved.
https://access.redhat.com/errata/RHBA-2021:0439