CUDACOMPILE : nvcc error : 'cudafe++' died with status 0xC0000005 (ACCESS_VIOLATION)

Accelerated Computing CUDA CUDA NVCC Compiler
1

Hi,

I just wrote the following (broken) code in a .cuh file:

template <typename Tv>
struct RowIndexer
{
  size_t operator()(size_t row, size_t col, size_t n) const
  {
    return row * n + col;
  }
};

template <typename Tv>
void bug()
{
  detail::RowIndexer rows;
}

Notice the absence of a template parameter to RowIndexer in the bug function.

Then, in a .cu file, I’m calling bug<float>();

This leads to the following error: CUDACOMPILE : nvcc error : 'cudafe++' died with status 0xC0000005 (ACCESS_VIOLATION)

There is no other indication that the code is incorrect. If I declare the bug function as a non-template function, the compiler catches the error.

This is all on MSVC 2022.