Using std:: integer types in device code

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi!

Quick question - are there any issues/concerns/implications with using C++ header <cstdint> and using std:: integer types, e.g. std::int32_t?

Example:

#include <cstdint>

__global__ void foo(std::int32_t x)
{
     ...
}

The code compiles and works (C++14), but I wonder if this is “correct” or can this cause issues in the future.

Thanks!

2

These types usually are just typedefs and work perfectly fine. If this worries you, you can use the official libcudacxx
GitHub - NVIDIA/libcudacxx: The C++ Standard Library for your entire system. which is a host/device standard library.

#include <cuda/std/cstdint>

__global__ void foo(cuda::std::int32_t* x){
    *x = 42;
}
3

Thanks for the info!

Unfortunately we are restricted to not using libcudaxx at the moment, so we need to go for the Standard headers.

Will this also work when cross-compiling for another target, say Jetson? Are the host typedefs guaranteed to be the same as the target typedefs?

4

After some more research, I conclude that:

  • Host and device must have the same size of ints, otherwise memcopying between host and device would not make sense.
  • NVCC will use whichever host compiler (x86 or aarch64) as preprocessor.
  • The preprocessed file will include the int types from /usr/include (x86) or /path/to/cross-toolchain/usr/include

So my conclusion is that host and device code use the same integer types, therefore it’s safe to use std::int32_t in device code.

Let me know if you see any flaw in my reasoning!