RadixSort does not compile for sm_20 in cub 1.8.0

[RaulPPelaez]

RadixSort tries to use shfl instructions even for arch <30, resulting in nvcc refusing to compile with these errors:

ptxas /tmp/tmpxft_00007c9d_00000000-6_ex.ptx, line 216; error   : Instruction 'shfl' requires .target sm_30 or higher
ptxas /tmp/tmpxft_00007c9d_00000000-6_ex.ptx, line 4240; error   : Instruction 'shfl' requires .target sm_30 or higher
ptxas /tmp/tmpxft_00007c9d_00000000-6_ex.ptx, line 5130; error   : Instruction 'shfl' requires .target sm_30 or higher
ptxas /tmp/tmpxft_00007c9d_00000000-6_ex.ptx, line 6751; error   : Instruction 'shfl' requires .target sm_30 or higher
ptxas /tmp/tmpxft_00007c9d_00000000-6_ex.ptx, line 7598; error   : Instruction 'shfl' requires .target sm_30 or higher
ptxas fatal   : Ptx assembly aborted due to errors

The simple code will not compile for me when using arch 20 in CUDA 8 or 7.5 (CUDA 9+ does not support arch 20 anyway...) .

#include"cub/cub.cuh"
int main(){
 size_t temp = 0;
int *ptr = nullptr;
cub::DeviceRadixSort::SortPairs(ptr, temp,
                                ptr, ptr, ptr, ptr, 0);
  return 0;

Compiling with nvcc -arch=sm_20 ex.cu will result in the errors above, while compiling with nvcc -srch=sm_30 ex.cu gives no error and compiles fine (do not try to run it though...).

[alliepiper]

sm20 is no longer supported. Policies for this architecture will be completely removed by #213.