[HIPIFY] hipify-clang terminates with constexpr function cannot overload function with same signature about hipify HOT 9 OPEN

The problem is that hipify-clang is awaiting clang header files in the 17 subfolder (in the newest format), whereas the rocm has an older format of this clang's subfolder: 17.0.0:

ignoring nonexistent directory "/opt/rocm/llvm/lib/clang/17/include/cuda_wrappers"
ignoring nonexistent directory "/opt/rocm/llvm/lib/clang/17/include"

What version of the ROCm and HIPIFY packages did you install?
Do they have the same version?

As an ugly workaround to make it work in your current configuration, you may copy the contains of the /opt/rocm/llvm/lib/clang/17.0.0/include/ to newly created /opt/rocm/llvm/lib/clang/17/include/. Explicit includes to clang like -I /opt/rocm/llvm/lib/clang/17/include/ are not needed, moreover, they can't help, as instead multiple -internal-isystem includes are used internally by hipify-clang itself. That is why you are getting multiple ignoring nonexistent directory messages in the verbose mode.

from hipify.

I am in arch installing both packages from extra repository. Current package version is (both):
5.7.1-1

If I symlink 17.0.0 to 17 in /opt/rocm/llvm/lib/clang/, I seem to be able to convert a file with the following command:

/opt/rocm/hip/bin/hipify-clang -v --inplace common.cu --cuda-path=/opt/cuda -I /opt/rocm/include/

from hipify.

What happened without -I /opt/rocm/llvm/lib/clang/17.0.0/include/?

from hipify.

Thanks for your answer. Can't test before January since I am traveling over Christmas 🎄. Will try and report back as soon as I am back in office!

from hipify.

Without the Include parameter, I am getting the following error:

<built-in>:1:10: fatal error: '__clang_cuda_runtime_wrapper.h' file not found
#include "__clang_cuda_runtime_wrapper.h"

from hipify.

Moving on to a more complex file, however, seems to greate more issues:

/tmp/aes_cbc.cu-6884ee.hip:199:16: error: redefinition of 'l_te' with a different type: 'const u32 *' (aka 'const unsigned int *') vs 'u32[256]' (aka 'unsigned int[256]')
    const u32* l_te = g_Te0;
               ^
/tmp/aes_cbc.cu-6884ee.hip:194:20: note: previous definition is here
    __shared__ u32 l_te[256];
                   ^
7 warnings and 1 error generated when compiling for host.
Error while processing /tmp/aes_cbc.cu-6884ee.hip.

I have to admit, that I have no / little experience with cuda and GPU programming at all, so this might take more time to investigate (If it even makes sense what I am doing) :/

Thanks for your help anyways! :)

from hipify.

Moving on to a more complex file, however, seems to greate more issues:

/tmp/aes_cbc.cu-6884ee.hip:199:16: error: redefinition of 'l_te' with a different type: 'const u32 *' (aka 'const unsigned int *') vs 'u32[256]' (aka 'unsigned int[256]')
    const u32* l_te = g_Te0;
               ^
/tmp/aes_cbc.cu-6884ee.hip:194:20: note: previous definition is here
    __shared__ u32 l_te[256];
                   ^
7 warnings and 1 error generated when compiling for host.
Error while processing /tmp/aes_cbc.cu-6884ee.hip.

I have to admit, that I have no / little experience with cuda and GPU programming at all, so this might take more time to investigate (If it even makes sense what I am doing) :/

Thanks for your help anyways! :)

Well, the compiler says that you have a redefinition of u32, and it looks like it is redefined. Have you tried to compile your CUDA source by CUDA's nvcc? Is the source correct?

from hipify.