"""Load metakernel natively on this machine (Windows/SAURON included).

The published Hugging Face kernel carries `*-linux` variants only, so on
Windows `get_kernel` has no matching build. The source is right here:
JIT-build it with `torch.utils.cpp_extension` and expose the identical
API.

    import load_local
    mk = load_local.load()
"""
import os
import sys
import types

_cached = None

_OPS = ["mk_triad", "mk_triad_passes", "mk_read", "mk_write", "mk_gather",
        "mk_fma_f64", "mk_fma_dep", "mk_chase_global", "mk_chase_shared",
        "mk_fma_f32", "mk_mma", "mk_atomics", "mk_spin", "mk_empty",
        "mk_barrier", "mk_occupancy"]


def load(verbose=False):
    """Return the metakernel module, JIT-built from the local source."""
    global _cached
    if _cached is not None:
        return _cached
    from torch.utils.cpp_extension import load as _jit

    root = os.path.dirname(os.path.abspath(__file__))
    cuda_dir = os.path.join(root, "metakernel_cuda")
    ext = _jit(name="metakernel_ext",
               sources=[os.path.join(root, "local_bind.cpp"),
                        os.path.join(cuda_dir, "probes.cu")],
               extra_include_paths=[cuda_dir],
               verbose=verbose)
    ops_mod = types.ModuleType("metakernel._ops")
    ops_mod.ops = type("_Ops", (), {
        name: staticmethod(getattr(ext, name)) for name in _OPS
    })()
    sys.modules["metakernel._ops"] = ops_mod
    sys.path.insert(0, os.path.join(root, "torch-ext"))
    import metakernel
    _cached = metakernel
    return metakernel