Issue compiling dgemm

[ashesh2512]

Hello, I have recently been exploring the tutorial to dive into HIP, and ran into the following issue when trying to compile the dgemm example.

make[2]: *** No rule to make target '/opt/rocm-5.3.0/hip/hipblas/lib/libhipblas.so', needed by 'dgemm'.  Stop.
make[1]: *** [CMakeFiles/Makefile2:126: src/CMakeFiles/dgemm.dir/all] Error 2
make: *** [Makefile:91: all] Error 2

I have the following modules loaded -

$ module list

Currently Loaded Modules:
  1) craype/2.7.19      3) libfabric/1.15.2.0   5) cray-libsci/22.12.1.1   7) hsi/default       9) amd/5.3.0          11) cmake/3.23.2
  2) cray-dsmml/0.2.2   4) craype-network-ofi   6) PrgEnv-amd/8.3.3        8) DefApps/default  10) cray-mpich/8.1.23

After some debugging, HIP_PATH is picking up the wrong path, in my case /opt/rocm-5.3.0/hip

After setting HIP_PATH to ROCM_PATH, I now run into an error in the linking stage -

[100%] Linking HIP executable ../dgemm
ld.lld: error: undefined symbol: pthread_create
>>> referenced by thread:126 (/usr/lib64/gcc/x86_64-suse-linux/7/../../../../include/c++/7/thread:126)
>>>               CMakeFiles/dgemm.dir/main.cpp.o:(std::__future_base::_Async_state_impl<std::thread::_Invoker<std::tuple<dgemm_results (*)(matrix<double> const&, matrix<double> const&, int, int, int), matrix<double>, matrix<double>, int, int, int>>, dgemm_results>::_Async_state_impl(std::thread::_Invoker<std::tuple<dgemm_results (*)(matrix<double> const&, matrix<double> const&, int, int, int), matrix<double>, matrix<double>, int, int, int>>&&))
clang-15: error: linker command failed with exit code 1 (use -v to see invocation)
make[2]: *** [src/CMakeFiles/dgemm.dir/build.make:144: dgemm] Error 1
make[1]: *** [CMakeFiles/Makefile2:126: src/CMakeFiles/dgemm.dir/all] Error 2
make: *** [Makefile:91: all] Error 2

What could I be still missing? Was it wrong to set the HIP_PATH to ROCM_PATH?

[tom-papatheodore]

I don't have access to any Cray systems, but I can at least comment on the environment variables. HIP_PATH should not be set to ROCM_PATH. ROCM_PATH should point to the main ROCm directory (e.g., /opt/rocm-x.x.x) and HIP_PATH should point to the hip directory within the ROCm directory (e.g., /opt/rocm-x.x.x/hip).

Things might be different when using the amd module on the specific Cray system you're running on. Which system is this out of curiosity?