Exception Handling

Standard C++ exception handling (try, catch, throw) is not supported in CUDA device code, while it is enabled by default in host code.

Device code

libcu++ maps exceptions to cuda::std::terminate() calls in device code, which translates to __trap() and terminates the kernel.

Host code

libcu++ allows users to manually disable exceptions in host code in two ways:

• By defining CCCL_DISABLE_EXCEPTIONS before including any library headers.

• By compiling with -fno-exceptions compiler flag with gcc or clang, or /EH- compiler flag with msvc.

If exceptions are disabled, a throw exception is translated into a cuda::std::terminate() call, which terminates the program.

cuda::cuda_error

Exception class thrown when a CUDA error is encountered. It inherits from std::runtime_error.

class cuda_error : public std::runtime_error
{
public:
    cuda_error(cudaError_t status, const char* msg);

    cudaError_t status() const noexcept;
};