conflicting types with cuda_runtime.h

[sthibaul]

Hello,

I'd like to use HIP to drive an AMD GPU, while using CUDA to drive a CUDA GPU. But the headers conflict

Samuel

$ cat test.c
#include <cuda_runtime.h>
#include <hip/hip_runtime.h>
int main(void) {
}
$ hipconfig -C
 -D__HIP_PLATFORM_HCC__= -D__HIP_PLATFORM_AMD__= -I/opt/rocm-5.1.0/hip/include -I/opt/rocm-5.1.0/llvm/bin/../lib/clang/14.0.0 -I/opt/rocm-5.1.0/hsa/include
$ gcc test.c -o test $(hipconfig -C)
In file included from /opt/rocm-5.1.0/hip/include/hip/amd_detail/amd_channel_descriptor.h:28,
                 from /opt/rocm-5.1.0/hip/include/hip/channel_descriptor.h:32,
                 from /opt/rocm-5.1.0/hip/include/hip/texture_types.h:38,
                 from /opt/rocm-5.1.0/hip/include/hip/hip_runtime_api.h:474,
                 from /opt/rocm-5.1.0/hip/include/hip/hip_runtime.h:113,
                 from test.c:2:
/opt/rocm-5.1.0/hip/include/hip/amd_detail/amd_hip_vector_types.h:1388:22: error: conflicting types for ‘uchar1’
 1388 | __MAKE_VECTOR_TYPE__(uchar, unsigned char);
      |                      ^~~~~
/opt/rocm-5.1.0/hip/include/hip/amd_detail/amd_hip_vector_types.h:1370:11: note: in definition of macro ‘__MAKE_VECTOR_TYPE__’
 1370 |         } CUDA_name##1;\
      |           ^~~~~~~~~
In file included from /usr/include/driver_types.h:59,
                 from /usr/include/builtin_types.h:59,
                 from /usr/include/cuda_runtime.h:91,
                 from test.c:1:
/usr/include/vector_types.h:362:42: note: previous declaration of ‘uchar1’ was here
  362 | typedef __device_builtin__ struct uchar1 uchar1;
      |                                          ^~~~~~
In file included from /opt/rocm-5.1.0/hip/include/hip/amd_detail/amd_channel_descriptor.h:28,
                 from /opt/rocm-5.1.0/hip/include/hip/channel_descriptor.h:32,
                 from /opt/rocm-5.1.0/hip/include/hip/texture_types.h:38,
                 from /opt/rocm-5.1.0/hip/include/hip/hip_runtime_api.h:474,
                 from /opt/rocm-5.1.0/hip/include/hip/hip_runtime.h:113,
                 from test.c:2:
[... same for uchar2, uchar3, ..., ‘struct textureReference’, ‘struct surfaceReference’, ‘struct dim3’, dim3]

[sthibaul]

for information, the reason why we need to include both cuda_runtime.h and hip_runtime.h is that in our runtime layer we have a data transfer manager and it wants to manage both cudaStream_t and hipStream_t in the same list of transfers to be performed

[SarbojitAMD]

@sthibaul, I would like to understand your development environment a little more.

Are you trying to run some tasks on AMD GPU and some on Nvidia GPU simultaneously? Hence you need both header files to be present?

[sthibaul]

Yes, that is it.

[ppanchad-amd]

@sthibaul Apologies for the lack of response. Can you please test with latest ROCm 6.0.2 (HIP 6.0.32831)? If resolved, please close ticket. Thanks!

[sthibaul]

I don't have 6.0.2 readily available, but 6.0.0 (1.0.0.60000-91~20.04 from https://repo.radeon.com/rocm/apt/debian) still has the issue.

[fwyzard]

Hi, I'm also interested in this issue, for the same reason: building an application that can offload operations to both NVIDIA and AMD GPUs.

The only way we found to do this is to split the CUDA-related and HIP/ROCm-related code in different source files, compile them separately, and link (statically or dynamically) the object files.

If there were a way to include both runtimes in the same source file I would also be very interested to use it!