Check error code from cudaOccupancyMaxActiveBlocksPerMultiprocessor

[malphil]

Hey all,

It's possible inside Kokkos_Cuda_Parallel.hpp for block.y = 0 in the following code, which can show up as an arithmetic divide-by-zero error.

  inline
  void execute() const
    {
      const typename Policy::index_type nwork = m_policy.end() - m_policy.begin();
      const int block_size = Kokkos::Impl::cuda_get_opt_block_size< ParallelFor, LaunchBounds>( m_functor , 1, 0 , 0 );
      const dim3 block(  1 , block_size , 1);
      const dim3 grid( std::min( typename Policy::index_type(( nwork + block.y - 1 ) / block.y) , typename Policy::index_type(cuda_internal_maximum_grid_count()) ) , 1 , 1);

      CudaParallelLaunch< ParallelFor, LaunchBounds >( *this , grid , block , 0 );
    }

Is it possible to check the error code from cudaOccupancyMaxActiveBlocksPerMultiprocessor or verify that cuda_get_opt_block_size does not return 0 prior to the division?

Thanks!