[SYCL][PI][HIP][CMake] Support non-standard ROCM installation paths

[ajaypanyala]

Is your feature request related to a problem? Please describe
The HIP build of the compiler fails in the configure step when ROCM is installed in a non-standard location especially when multiple rocm versions exist and are loaded via modules. The CMake infrastructure for the HIP plugins assumes it to always be in /opt/rocm as shown here and here

The configure step used is as follows:
CC=gcc CXX=g++ python ./buildbot/configure.py -w ./build_rocm --hip --hip-platform AMD -o ./

Describe the solution you would like
This is already supported in other places as shown here. Similar ROCM discovery logic needs to be added to the HIP plugins part.

[npmiller]

You're right that we could improve ROCm installation detection but note that you can already point the build to a different ROCm directory, for example, if say your ROCm installation is in /usr/local/rocm:

CC=gcc CXX=g++ python ./buildbot/configure.py -w ./build_rocm --hip --hip-platform AMD -o . \
  --cmake-opt=-DSYCL_BUILD_PI_HIP_INCLUDE_DIR=/usr/local/rocm/hip/include \
  --cmake-opt=-DSYCL_BUILD_PI_HIP_HSA_INCLUDE_DIR=/usr/local/rocm/hsa/include \
  --cmake-opt=-DSYCL_BUILD_PI_HIP_AMD_LIBRARY=/usr/local/rocm/hip/lib/libamdhip64.so

[ajaypanyala]

@npmiller Thanks! I added the additional cmake-options you suggested to the configure line and it works well!

[zjin-lcf]

I hope the solution may be added to the getting started guide.

[npmiller]

The variables are already described in the getting started guide:

• https://intel.github.io/llvm-docs/GetStartedGuide.html#build-dpc-toolchain-with-support-for-hip-amd

I can add a note that they can be specified with --cmake-opt if you want

[zjin-lcf]

The note will be helpful. Thanks.