Fix CUDA 9 sm variables.

[kazuki0824]

"compute_72" is not supported in CUDA 9. See also https://docs.nvidia.com/cuda/archive/9.0/cuda-compiler-driver-nvcc/index.html#virtual-architecture-feature-list.

[taketwo]

Sorry I don't see any proof that 72 is not supported in CUDA 9 on the page you referenced. Contrary, Wikipedia claims that

CUDA SDK 9.0/9.1/9.2 support for compute capability 3.0 – 7.2 (Kepler, Maxwell, Pascal, Volta)

[kazuki0824]

1. NVCC emits the error below in makefile.

[ 58%] Building NVCC (Device) object cuda/segmentation/CMakeFiles/pcl_cuda_segmentation.dir/src/pcl_cuda_segmentation_generated_connected_components.cu.o
nvcc fatal : Unsupported gpu architecture 'compute_72'
CMake Error at pcl_cuda_segmentation_generated_connected_components.cu.o.cmake:209 (message):
Error generating
/pcl/release/cuda/segmentation/CMakeFiles/pcl_cuda_segmentation.dir/src/./pcl_cuda_segmentation_generated_connected_components.cu.o

make[2]: *** [cuda/segmentation/CMakeFiles/pcl_cuda_segmentation.dir/src/pcl_cuda_segmentation_generated_connected_components.cu.o] Error 1

Environment: Inside Docker, Ubuntu 16.04 + CUDA 9.0
CMake arguments: cmake -DCMAKE_BUILD_TYPE=None -DBUILD_CUDA=ON -DBUILD_GPU=ON -WITH_CUDA=ON -DBUILD_common=ON -DBUILD_features=ON -DBUILD_filters=OFF -DBUILD_global_tests=OFF -DBUILD_io=ON -DBUILD_kdtree=ON -DBUILD_keypoints=OFF -DBUILD_stereo=OFF -DBUILD_registration=ON -DBUILD_sample_consensus=OFF -DBUILD_segmentation=OFF -DBUILD_surface=OFF -DWITH_DOCS=OFF -DWITH_TUTORIALS=OFF -DCMAKE_INSTALL_PREFIX=/usr ..

2. nvcc --help actually says below.
   nvcc --help in CUDA 9.1
   Allowed values for this option: 'compute_30','compute_32','compute_35',
   'compute_37','compute_50','compute_52','compute_53','compute_60','compute_61',
   'compute_62','compute_70','compute_72','sm_30','sm_32','sm_35','sm_37','sm_50',
   'sm_52','sm_53','sm_60','sm_61','sm_62','sm_70','sm_72'.
   nvcc --help in CUDA 9.0
   Allowed values for this option: 'compute_30','compute_32','compute_35',
   'compute_37','compute_50','compute_52','compute_53','compute_60','compute_61',
   'compute_62','compute_70','sm_30','sm_32','sm_35','sm_37','sm_50','sm_52',
   'sm_53','sm_60','sm_61','sm_62','sm_70'.

Thus, CUDA 9.0 is incapable of dealing with sm_72. Please check it out also.

[taketwo]

Thanks for the details. Could you please update the PR such that both your problem is solved and CUDA 9.1 users can use sm_72? That is, add one more else branch.

[kazuki0824]

As @taketwo says, this topic has been moved to #2567.

[SergioRAgostinho]

1. nvcc --help actually says below.
   nvcc --help in CUDA 9.1
   Allowed values for this option:  'compute_30','compute_32','compute_35',
   'compute_37','compute_50','compute_52','compute_53','compute_60','compute_61',
   'compute_62','compute_70','compute_72','sm_30','sm_32','sm_35','sm_37','sm_50',
   'sm_52','sm_53','sm_60','sm_61','sm_62','sm_70','sm_72'.
   nvcc --help in CUDA 9.0
   Allowed values for this option:  'compute_30','compute_32','compute_35',
   'compute_37','compute_50','compute_52','compute_53','compute_60','compute_61',
   'compute_62','compute_70','sm_30','sm_32','sm_35','sm_37','sm_50','sm_52',
   'sm_53','sm_60','sm_61','sm_62','sm_70'.

Cool tip. Didn't know this one. Thanks