[QMCPack] Use QMC_GPU instead of removed ENABLE_CUDA/QMC_CUDA2HIP

[ro-i]

QMCPack removed ENABLE_CUDA, ENABLE_ROCM, QMC_CUDA2HIP, ENABLE_SYCL and ENABLE_OFFLOAD in b84e4563 (Jan 2025) and now hard-errors at CMakeLists.txt:71 if any of them is defined. Use QMC_GPU, the replacement documented in docs/installation.rst.

[jplehr]

Is offload-arch detected automatically?
If so: How does it handle multi-GPU systems?

[ro-i]

only now with QMC_GPU_ARCHS are we actually getting offload-arch:

    if(OFFLOAD_TARGET)
      set(OPENMP_OFFLOAD_COMPILE_OPTIONS "-fopenmp-targets=${OFFLOAD_TARGET}")
      if(OFFLOAD_ARCH)
        set(OPENMP_OFFLOAD_COMPILE_OPTIONS
            "${OPENMP_OFFLOAD_COMPILE_OPTIONS} -Xopenmp-target=${OFFLOAD_TARGET} -march=${OFFLOAD_ARCH}")
      endif()
    elseif(QMC_GPU_ARCHS)
      string(REGEX REPLACE ";" "," QMC_GPU_ARCHS_COMMA_SEPARATED "${QMC_GPU_ARCHS}")
      set(OPENMP_OFFLOAD_COMPILE_OPTIONS "--offload-arch=${QMC_GPU_ARCHS_COMMA_SEPARATED}")
    else()
      message(FATAL_ERROR "Require QMC_GPU_ARCHS or OFFLOAD_TARGET set for OpenMP offload using Clang.")
    endif()

https://github.com/QMCPACK/qmcpack/blob/61175a9ebaa4a4829061c03b398cd6baff9d95f8/CMake/ClangCompilers.cmake#L23-L34

[ro-i]

I don't think that any of this affects multi-GPU support (neither positive, nor negative)

[jplehr]

Ok, so we will simply build for all the archs found on a system?
I'm just asking because the auto-detection of the arch(s) has bitten us in the past. ;)

[ro-i]

hm, so previously, OpenMP got its arch via -DOFFLOAD_ARCH=$AOMP_GPU and HIP got its arches by qmcpack-side autodetection (which led to all arches on a given system).

Now, both OpenMP and HIP get their arch from $AOMP_GPU. And this is set in the aomp-side autodetection (which is limited to one arch at the moment - independent of this PR).

[jplehr]

Apologies, I did not see line 123 change / addition.

[ro-i]

@jplehr can you merge it, please? aomp repository permissions seem to have moved as well ._.