Add support for MacOS

[proski]

This is for x86_64 only. No attempt was made to support Apple Silicon. However, this PR could be used as a spepping stone in that direction.

The remaining problem is an error when running prpll:

Exception gpu_error: INVALID_WORK_GROUP_SIZE (-54) transpIn at /Users/plroskin/src/gpuowl/src/clwrap.cpp:312 run

That would need to be addressed by an OpenCL expert. It's possible that the issue is specific to my hardware or software.

However, many other issues have been addressed is a clean way.

• The real gcc (e.g. from homebrew) is selected for the build (clang doesn't support 128-bit floating point and std::filesystem::path in the standard library).
• OpenCL is linked the way is should be on MacOS with the -framework OpenCL linker directive.
• The cmake build uses cmake facilities to search for OpenCL and link to it.
• A replacement for fdatasync() is used on MacOS.
• OpenCL functions have prototypes when needed.
• Comparison between i32 and u32 in OpenCL code has been fixed.
• transposeWords uses signature that preserves global and restrict attributes used by the callers.
• clCreateCommandQueueWithProperties is missing on MacOS, use clCreateCommandQueueWithPropertiesAPPLE instead.

Basically, the C++ code compiles, the OpenCL code compiles and loads, but fails to run on the GPU.

[tdulcet]

@proski - If you want future commits to be automatically checked that they successfully build on x86 macOS, you could add something like this to the existing GitHub actions workflow file:

  macOS:
    name: macOS

    runs-on: macos-13
    steps:
    - uses: actions/checkout@v4
    - name: Install
      run: |
        brew install opencl-headers
        g++ --version
    - name: Script
      run: |
        make prpll -j "$(sysctl -n hw.ncpu)"
        cd build-release
        rm -f -- *.o
        ./prpll -h
    - uses: actions/upload-artifact@v4
      if: always()
      with:
        name: macos_prpll
        path: ${{ github.workspace }}

I am guessing that the opencl-headers package would need to be installed. The opencl-icd-loader package might need to be installed as well.

[proski]

@proski - If you want future commits to be automatically checked that they successfully build on x86 macOS, you could add something like this to the existing GitHub actions workflow file:

I've added it, thank you!

I am guessing that the opencl-headers package would need to be installed. The opencl-icd-loader package might need to be installed as well.

None of them is needed. Even gcc@14 is installed already, but I'm installing it anyway to be explicit about the dependency.

We should run a real test like prpll -iters 10000 -ll 82589933 but other OSes don't do it, so let's keep it consistent. Also, there is no test for the cmake build, I should have added it in #293. It could be added in a separate PR.

[proski]

INVALID_WORK_GROUP_SIZE is reproducible in the CI:

https://github.com/preda/gpuowl/actions/runs/10842867463/job/30089143591?pr=295#step:4:71

[proski]

Added fallback for the case when cmake cannot find OpenCL but linking with -lOpenCL works. Observed on Windows in the CI.