__CUDA_ARCH__ is defined when compiling AOT for HIP

[AuroraPerego]

Describe the bug

During ahead-of-time compilation for the HIP backend, the macro __CUDA_ARCH__ is defined as 0, while it should not be defined at all. Usually, libraries check only if __CUDA_ARCH__ is defined to enable CUDA specific code, without checking its value, causing compilation for the HIP backend to fail.

To reproduce

1. Include code snippet as short as possible

test.cpp

#include <sycl/sycl.hpp>
int main()
{
    return 0;
}

2. Specify the command which should be used to compile the program

icpx -fsycl -fsycl-targets=amd_gpu_gfx90a -dM -E test.cpp | grep CUDA_ARCH

3. Indicate what is wrong and what was expected
   the output of the command above is

#define __CUDA_ARCH__ 0

while it should not be defined.

Environment

• OS: RHEL 8.10
• Target device and vendor: AMD GPU MI250X
• icpx version:

  Intel(R) oneAPI DPC++/C++ Compiler 2024.2.1 (2024.2.1.20240711)
  Target: x86_64-unknown-linux-gnu
  Thread model: posix
  InstalledDir: /opt/intel/oneapi/compiler/2024.2/bin/compiler
  Configuration file: /opt/intel/oneapi/compiler/2024.2/bin/compiler/../icpx.cfg
  

• Dependencies version:

  HIP version: 6.2.41133-dd7f95766
  AMD clang version 18.0.0git (https://github.com/RadeonOpenCompute/llvm-project roc-6.2.0 24292 26466ce804ac523b398608f17388eb6d605a3f09)
  Target: x86_64-unknown-linux-gnu
  Thread model: posix
  InstalledDir: /opt/rocm-6.2.0/lib/llvm/bin
  Configuration file: /opt/rocm-6.2.0/lib/llvm/bin/clang++.cfg
  

Additional context

No response

[AuroraPerego]

FYI @fwyzard @ivorobts

[GeorgeWeb]

Hi, good timing with this and good catch. I also noticed that happening while fixing another regression defining this macro in the SYCL for Cuda AOT compilation (where we want 'SYCL_CUDA_ARCH' instead) and intended to follow up with a HIP patch-fix afterwards.
I'll ping you when it's up. :)

[npmiller]

This should be fixed by: #15441

Nevermind I got my patches mixed up, the patches for this isn't up yet