[meta_schedule]Try to use cuda-tensorcore in meta schedule, but get not registered error

junrushao · April 10, 2023, 4:17am

We will need to import this module to register intrinsics:

import tvm.tir.tensor_intrin  # pylint: disable=unused-import

It’s usually lazily imported when creating MetaSchedule tuning contexts, but for some reason it’s not triggered in your case

github.com

apache/tvm/blob/17bd178bfad8d831e983cd91d76ca29f1f32e422/python/tvm/meta_schedule/tune_context.py#L108


    *,
    target: Union[Target, str, None] = None,
    space_generator: Union["SpaceGenerator.SpaceGeneratorType", None] = None,
    search_strategy: Union["SearchStrategy.SearchStrategyType", None] = None,
    task_name: str = "main",
    rand_state: int = -1,
    num_threads: Union[int, Literal["physical", "logical"]] = "physical",
    logger: Optional[Logger] = None,
):
    # pylint: disable=import-outside-toplevel
    import tvm.tir.tensor_intrin  # pylint: disable=unused-import


    from .search_strategy import SearchStrategy
    from .space_generator import SpaceGenerator


    # pylint: enable=import-outside-toplevel
    if isinstance(mod, PrimFunc):
        mod = _normalize_mod(mod)
    if target is not None:
        if not isinstance(target, Target):
            target = Target(target)