Failed to build against cudnn 8.0.2

[jiapei100]

System information (version)

• OpenCV => 4.4
• Operating System / Platform => Ubuntu 20.04
• Compiler => gcc/g++ 9.3.0
• Cuda => 11.0
• NVidia Driver => 450.57
• GPU: Geforce 2080 Ti

Detailed description

[ 32%] Building NVCC (Device) object modules/dnn/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_detection_output.cu.o
cd ....../opencv/build/modules/dnn/CMakeFiles/cuda_compile_1.dir/src/cuda && /snap/cmake/513/bin/cmake -E make_directory ....../opencv/build/modules/dnn/CMakeFiles/cuda_compile_1.dir/src/cuda/.
cd ....../opencv/build/modules/dnn/CMakeFiles/cuda_compile_1.dir/src/cuda && /snap/cmake/513/bin/cmake -D verbose:BOOL=ON -D build_configuration:STRING=Release -D generated_file:STRING=....../opencv/build/modules/dnn/CMakeFiles/cuda_compile_1.dir/src/cuda/./cuda_compile_1_generated_detection_output.cu.o -D generated_cubin_file:STRING=....../opencv/build/modules/dnn/CMakeFiles/cuda_compile_1.dir/src/cuda/./cuda_compile_1_generated_detection_output.cu.o.cubin.txt -P ....../opencv/build/modules/dnn/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_detection_output.cu.o.Release.cmake
-- Removing ....../opencv/build/modules/dnn/CMakeFiles/cuda_compile_1.dir/src/cuda/./cuda_compile_1_generated_detection_output.cu.o
/snap/cmake/513/bin/cmake -E rm -f ....../opencv/build/modules/dnn/CMakeFiles/cuda_compile_1.dir/src/cuda/./cuda_compile_1_generated_detection_output.cu.o
-- Generating dependency file: ....../opencv/build/modules/dnn/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_detection_output.cu.o.NVCC-depend
/usr/local/cuda/bin/nvcc -M -D__CUDACC__ ....../opencv/modules/dnn/src/cuda/detection_output.cu -o ....../opencv/build/modules/dnn/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_detection_output.cu.o.NVCC-depend -ccbin /usr/bin/cc -m64 -DVK_NO_PROTOTYPES -D__OPENCV_BUILD=1 -D_USE_MATH_DEFINES -D__STDC_CONSTANT_MACROS -D__STDC_LIMIT_MACROS -D__STDC_FORMAT_MACROS -DCV_CUDA4DNN=1 -DOPENCV_DNN_EXTERNAL_PROTOBUF=1 -DHAVE_PROTOBUF=1 -Xcompiler ,\"-fsigned-char\",\"-ffast-math\",\"-W\",\"-Wall\",\"-Werror=return-type\",\"-Werror=non-virtual-dtor\",\"-Werror=address\",\"-Werror=sequence-point\",\"-Wformat\",\"-Werror=format-security\",\"-Winit-self\",\"-Wpointer-arith\",\"-Wuninitialized\",\"-Winit-self\",\"-Wno-comment\",\"-Wno-strict-overflow\",\"-fdiagnostics-show-option\",\"-Wno-long-long\",\"-pthread\",\"-fno-omit-frame-pointer\",\"-pg\",\"-g\",\"-msse\",\"-msse2\",\"-msse3\",\"-fvisibility=hidden\",\"-fopenmp\",\"-Wno-deprecated\",\"-Wno-missing-declarations\",\"-Wno-shadow\",\"-Wno-unused-parameter\",\"-Wno-sign-compare\",\"-Wno-undef\",\"-Wno-invalid-offsetof\",\"-Wno-unused-but-set-variable\",\"-O3\",\"-DNDEBUG\",\"-DNDEBUG\" -gencode arch=compute_75,code=sm_75 -D_FORCE_INLINES --use_fast_math -Xcompiler -DCVAPI_EXPORTS -Xcompiler -fPIC --std=c++14 -DNVCC -I/usr/local/cuda/include -I....../opencv/build -I/usr/include/va -I/usr/include/eigen3/Eigen -I/usr/include/x86_64-linux-gnu -I....../opencv/3rdparty/include -I....../opencv/modules/dnn/include -I....../opencv/build/modules/dnn -I....../opencv_contrib/modules/cudev/include -I....../opencv/modules/core/include -I....../opencv/modules/imgproc/include -I/usr/local/include
-- Generating temporary cmake readable file: ....../opencv/build/modules/dnn/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_detection_output.cu.o.depend.tmp
/snap/cmake/513/bin/cmake -D input_file:FILEPATH=....../opencv/build/modules/dnn/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_detection_output.cu.o.NVCC-depend -D output_file:FILEPATH=....../opencv/build/modules/dnn/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_detection_output.cu.o.depend.tmp -D verbose=ON -P /snap/cmake/513/share/cmake-3.18/Modules/FindCUDA/make2cmake.cmake
-- Copy if different ....../opencv/build/modules/dnn/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_detection_output.cu.o.depend.tmp to ....../opencv/build/modules/dnn/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_detection_output.cu.o.depend
/snap/cmake/513/bin/cmake -E copy_if_different ....../opencv/build/modules/dnn/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_detection_output.cu.o.depend.tmp ....../opencv/build/modules/dnn/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_detection_output.cu.o.depend
-- Removing ....../opencv/build/modules/dnn/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_detection_output.cu.o.depend.tmp and ....../opencv/build/modules/dnn/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_detection_output.cu.o.NVCC-depend
/snap/cmake/513/bin/cmake -E rm -f ....../opencv/build/modules/dnn/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_detection_output.cu.o.depend.tmp ....../opencv/build/modules/dnn/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_detection_output.cu.o.NVCC-depend
-- Generating ....../opencv/build/modules/dnn/CMakeFiles/cuda_compile_1.dir/src/cuda/./cuda_compile_1_generated_detection_output.cu.o
/usr/local/cuda/bin/nvcc ....../opencv/modules/dnn/src/cuda/detection_output.cu -c -o ....../opencv/build/modules/dnn/CMakeFiles/cuda_compile_1.dir/src/cuda/./cuda_compile_1_generated_detection_output.cu.o -ccbin /usr/bin/cc -m64 -DVK_NO_PROTOTYPES -D__OPENCV_BUILD=1 -D_USE_MATH_DEFINES -D__STDC_CONSTANT_MACROS -D__STDC_LIMIT_MACROS -D__STDC_FORMAT_MACROS -DCV_CUDA4DNN=1 -DOPENCV_DNN_EXTERNAL_PROTOBUF=1 -DHAVE_PROTOBUF=1 -Xcompiler ,\"-fsigned-char\",\"-ffast-math\",\"-W\",\"-Wall\",\"-Werror=return-type\",\"-Werror=non-virtual-dtor\",\"-Werror=address\",\"-Werror=sequence-point\",\"-Wformat\",\"-Werror=format-security\",\"-Winit-self\",\"-Wpointer-arith\",\"-Wuninitialized\",\"-Winit-self\",\"-Wno-comment\",\"-Wno-strict-overflow\",\"-fdiagnostics-show-option\",\"-Wno-long-long\",\"-pthread\",\"-fno-omit-frame-pointer\",\"-pg\",\"-g\",\"-msse\",\"-msse2\",\"-msse3\",\"-fvisibility=hidden\",\"-fopenmp\",\"-Wno-deprecated\",\"-Wno-missing-declarations\",\"-Wno-shadow\",\"-Wno-unused-parameter\",\"-Wno-sign-compare\",\"-Wno-undef\",\"-Wno-invalid-offsetof\",\"-Wno-unused-but-set-variable\",\"-O3\",\"-DNDEBUG\",\"-DNDEBUG\" -gencode arch=compute_75,code=sm_75 -D_FORCE_INLINES --use_fast_math -Xcompiler -DCVAPI_EXPORTS -Xcompiler -fPIC --std=c++14 -DNVCC -I/usr/local/cuda/include -I....../opencv/build -I/usr/include/va -I/usr/include/eigen3/Eigen -I/usr/include/x86_64-linux-gnu -I....../opencv/3rdparty/include -I....../opencv/modules/dnn/include -I....../opencv/build/modules/dnn -I....../opencv_contrib/modules/cudev/include -I....../opencv/modules/core/include -I....../opencv/modules/imgproc/include -I/usr/local/include
....../opencv/modules/dnn/src/cuda/detection_output.cu: In instantiation of ?typename std::enable_if<(current != 0), void>::type cv::dnn::cuda4dnn::kernels::dispatch_decode_bboxes(int, Args&& ...) [with T = __half; long unsigned int current = 15; Args = {const cv::dnn::cuda4dnn::csl::Stream&, cv::dnn::cuda4dnn::csl::Span<__half>&, cv::dnn::cuda4dnn::csl::Span<const __half>&, cv::dnn::cuda4dnn::csl::Span<const __half>&, bool&, bool&, long unsigned int&, long unsigned int&, float&, float&}; typename std::enable_if<(current != 0), void>::type = void]?:
....../opencv/modules/dnn/src/cuda/detection_output.cu:704:32:   required from ?void cv::dnn::cuda4dnn::kernels::decode_bboxes(const cv::dnn::cuda4dnn::csl::Stream&, cv::dnn::cuda4dnn::csl::Span<T>, cv::dnn::cuda4dnn::csl::View<T>, cv::dnn::cuda4dnn::csl::View<T>, std::size_t, bool, std::size_t, bool, bool, bool, bool, bool, float, float) [with T = __half; cv::dnn::cuda4dnn::csl::View<T> = cv::dnn::cuda4dnn::csl::Span<const __half>; std::size_t = long unsigned int]?
....../opencv/modules/dnn/src/cuda/detection_output.cu:707:391:   required from here
....../opencv/modules/dnn/src/cuda/detection_output.cu:687:92: error: no matching function for call to ?launch_decode_boxes_kernel<__half, (15 & 8), (15 & 4), (15 & 2), (15 & 1)>(const cv::dnn::cuda4dnn::csl::Stream&, cv::dnn::cuda4dnn::csl::Span<__half>&, cv::dnn::cuda4dnn::csl::Span<const __half>&, cv::dnn::cuda4dnn::csl::Span<const __half>&, bool&, bool&, long unsigned int&, long unsigned int&, float&, float&)?
  687 |         launch_decode_boxes_kernel<T, current & 8, current & 4, current & 2, current & 1>(std::forward<Args>(args)...);
      |         ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~   
....../opencv/modules/dnn/src/cuda/detection_output.cu:666:1: note: candidate: ?template<class T, bool SHARE_LOCATION, bool VARIANCE_ENCODED_IN_TARGET, bool CORNER_TRUE_CENTER_FALSE, bool CLIP_BBOX> void cv::dnn::cuda4dnn::kernels::launch_decode_boxes_kernel(const cv::dnn::cuda4dnn::csl::Stream&, cv::dnn::cuda4dnn::csl::Span<T>, cv::dnn::cuda4dnn::csl::View<T>, cv::dnn::cuda4dnn::csl::View<T>, bool, bool, cv::dnn::cuda4dnn::csl::device::size_type, cv::dnn::cuda4dnn::csl::device::index_type, float, float)?
  666 | void launch_decode_boxes_kernel(const Stream& stream, Span<T> decoded_bboxes, View<T> locations, View<T> priors,
      | ^~~~~~~~~~~~~~~~~~~~~~~~~~
....../opencv/modules/dnn/src/cuda/detection_output.cu:666:1: note:   template argument deduction/substitution failed:
....../opencv/modules/dnn/src/cuda/detection_output.cu:687:92: error: narrowing conversion of ?8? from ?long unsigned int? to ?bool? [-Wnarrowing]
  687 |         launch_decode_boxes_kernel<T, current & 8, current & 4, current & 2, current & 1>(std::forward<Args>(args)...);
      |         ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~   
....../opencv/modules/dnn/src/cuda/detection_output.cu:687:92: error: narrowing conversion of ?4? from ?long unsigned int? to ?bool? [-Wnarrowing]
....../opencv/modules/dnn/src/cuda/detection_output.cu:687:92: error: narrowing conversion of ?2? from ?long unsigned int? to ?bool? [-Wnarrowing]
....../opencv/modules/dnn/src/cuda/detection_output.cu: In instantiation of ?typename std::enable_if<(current != 0), void>::type cv::dnn::cuda4dnn::kernels::dispatch_decode_bboxes(int, Args&& ...) [with T = float; long unsigned int current = 15; Args = {const cv::dnn::cuda4dnn::csl::Stream&, cv::dnn::cuda4dnn::csl::Span<float>&, cv::dnn::cuda4dnn::csl::Span<const float>&, cv::dnn::cuda4dnn::csl::Span<const float>&, bool&, bool&, long unsigned int&, long unsigned int&, float&, float&}; typename std::enable_if<(current != 0), void>::type = void]?:
....../opencv/modules/dnn/src/cuda/detection_output.cu:704:32:   required from ?void cv::dnn::cuda4dnn::kernels::decode_bboxes(const cv::dnn::cuda4dnn::csl::Stream&, cv::dnn::cuda4dnn::csl::Span<T>, cv::dnn::cuda4dnn::csl::View<T>, cv::dnn::cuda4dnn::csl::View<T>, std::size_t, bool, std::size_t, bool, bool, bool, bool, bool, float, float) [with T = float; cv::dnn::cuda4dnn::csl::View<T> = cv::dnn::cuda4dnn::csl::Span<const float>; std::size_t = long unsigned int]?
....../opencv/modules/dnn/src/cuda/detection_output.cu:708:388:   required from here
....../opencv/modules/dnn/src/cuda/detection_output.cu:687:92: error: no matching function for call to ?launch_decode_boxes_kernel<float, (15 & 8), (15 & 4), (15 & 2), (15 & 1)>(const cv::dnn::cuda4dnn::csl::Stream&, cv::dnn::cuda4dnn::csl::Span<float>&, cv::dnn::cuda4dnn::csl::Span<const float>&, cv::dnn::cuda4dnn::csl::Span<const float>&, bool&, bool&, long unsigned int&, long unsigned int&, float&, float&)?
....../opencv/modules/dnn/src/cuda/detection_output.cu:666:1: note: candidate: ?template<class T, bool SHARE_LOCATION, bool VARIANCE_ENCODED_IN_TARGET, bool CORNER_TRUE_CENTER_FALSE, bool CLIP_BBOX> void cv::dnn::cuda4dnn::kernels::launch_decode_boxes_kernel(const cv::dnn::cuda4dnn::csl::Stream&, cv::dnn::cuda4dnn::csl::Span<T>, cv::dnn::cuda4dnn::csl::View<T>, cv::dnn::cuda4dnn::csl::View<T>, bool, bool, cv::dnn::cuda4dnn::csl::device::size_type, cv::dnn::cuda4dnn::csl::device::index_type, float, float)?
  666 | void launch_decode_boxes_kernel(const Stream& stream, Span<T> decoded_bboxes, View<T> locations, View<T> priors,
      | ^~~~~~~~~~~~~~~~~~~~~~~~~~
....../opencv/modules/dnn/src/cuda/detection_output.cu:666:1: note:   template argument deduction/substitution failed:
....../opencv/modules/dnn/src/cuda/detection_output.cu:687:92: error: narrowing conversion of ?8? from ?long unsigned int? to ?bool? [-Wnarrowing]
  687 |         launch_decode_boxes_kernel<T, current & 8, current & 4, current & 2, current & 1>(std::forward<Args>(args)...);
      |         ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~   
....../opencv/modules/dnn/src/cuda/detection_output.cu:687:92: error: narrowing conversion of ?4? from ?long unsigned int? to ?bool? [-Wnarrowing]
....../opencv/modules/dnn/src/cuda/detection_output.cu:687:92: error: narrowing conversion of ?2? from ?long unsigned int? to ?bool? [-Wnarrowing]
....../opencv/modules/dnn/src/cuda/detection_output.cu:685:1: warning: ?typename std::enable_if<(current != 0), void>::type cv::dnn::cuda4dnn::kernels::dispatch_decode_bboxes(int, Args&& ...) [with T = float; long unsigned int current = 14; Args = {const cv::dnn::cuda4dnn::csl::Stream&, cv::dnn::cuda4dnn::csl::Span<float>&, cv::dnn::cuda4dnn::csl::Span<const float>&, cv::dnn::cuda4dnn::csl::Span<const float>&, bool&, bool&, long unsigned int&, long unsigned int&, float&, float&}]? used but never defined
  685 | ::type dispatch_decode_bboxes(int selector, Args&& ...args) {
      | ^~~~~~~~~~~~~~~~~~~~~~
....../opencv/modules/dnn/src/cuda/detection_output.cu:685:1: warning: ?typename std::enable_if<(current != 0), void>::type cv::dnn::cuda4dnn::kernels::dispatch_decode_bboxes(int, Args&& ...) [with T = __half; long unsigned int current = 14; Args = {const cv::dnn::cuda4dnn::csl::Stream&, cv::dnn::cuda4dnn::csl::Span<__half>&, cv::dnn::cuda4dnn::csl::Span<const __half>&, cv::dnn::cuda4dnn::csl::Span<const __half>&, bool&, bool&, long unsigned int&, long unsigned int&, float&, float&}]? used but never defined
-- Removing ....../opencv/build/modules/dnn/CMakeFiles/cuda_compile_1.dir/src/cuda/./cuda_compile_1_generated_detection_output.cu.o
/snap/cmake/513/bin/cmake -E rm -f ....../opencv/build/modules/dnn/CMakeFiles/cuda_compile_1.dir/src/cuda/./cuda_compile_1_generated_detection_output.cu.o
CMake Error at cuda_compile_1_generated_detection_output.cu.o.Release.cmake:280 (message):
  Error generating file
  ....../opencv/build/modules/dnn/CMakeFiles/cuda_compile_1.dir/src/cuda/./cuda_compile_1_generated_detection_output.cu.o


make[2]: *** [modules/dnn/CMakeFiles/opencv_dnn.dir/build.make:3057: modules/dnn/CMakeFiles/cuda_compile_1.dir/src/cuda/cuda_compile_1_generated_detection_output.cu.o] Error 1
make[2]: Leaving directory '....../opencv/build'
make[1]: *** [CMakeFiles/Makefile2:4371: modules/dnn/CMakeFiles/opencv_dnn.dir/all] Error 2
make[1]: Leaving directory '....../opencv/build'
make: *** [Makefile:185: all] Error 2

[jiapei100]

Here we are....

A lot of people have compalined gcc 9.3. Therefore, I tried clang10, which successfully built opencv4.4.0-dev, after some trivial modification (Refer to #17952 ), and the file modified is gkernel.hpp

[YashasSamaga]

I have reproduced this problem. It's happening in gcc 9 only. (compiles without errors in gcc 8 and below).

The following patch fixes the problem.

- launch_decode_boxes_kernel<T, current & 8, current & 4, current & 2, current & 1>(std::forward<Args>(args)...); 
+ launch_decode_boxes_kernel<T,
                                   static_cast<bool>(current & 8),
                                   static_cast<bool>(current & 4),
                                   static_cast<bool>(current & 2),
                                   static_cast<bool>(current & 1)
                                  >(std::forward<Args>(args)...);

[YashasSamaga]

Therefore, I tried clang10, which successfully built opencv4.4.0-dev, after some trivial modification (Refer to #17952 ), and the file modified is gkernel.hpp

clang10 is throwing the same error.

MCVE: https://godbolt.org/

MSVC and gcc 8 and below have no problem but all clang versions have a problem. clang is much more pedantic than gcc.

[jiapei100]

@YashasSamaga

Hi, your solution must be correct.
I successfully built OpenCV-4.4.0 Release yesterday. It seems current git version needs your solution...
I just release my clang-10 successfully compiled a version of OpenCL instead of CUDA.

[YashasSamaga]

I think this issue should stay open. It's a bug. gcc 9 and clang are correct about the error.