[FEA]: Generalize autotuning to dynamically generated kernels

まず、issue で説明されているエントリーポイント exhaustive_search と _TimingCandidate の使用箇所を読み、次に、現在どのように設定からカーネル、グリッド、ヒント、引数が生成されているかを追跡します。各設定から生成されたカーネルと引数を exhaustive_search がサポートでき、既存の固定カーネルの動作と autotuning の候補を維持できれば、作業は完了です。

Is this a new feature, an improvement, or a change to existing functionality?

Improvement

How would you describe the priority of this feature request?

Low (would be nice)

Please provide a clear description of problem this feature solves

Currently, exhaustive_search takes a single fixed kernel and construct arguments from an arbitrary sequence of configurations. That it takes a fixed kernel makes it unusable in its current form if the config objects themselves generates the kernel.

Feature Description

I'm currently using cutile and metaprogramming to generate kernels, with much better success and less pain than other frameworks. However, I can't use exhaustive_search in its current form, but need to modify it to take kernel generating functions.

Describe your ideal solution

The proposal is essentially to change exhaustive_search, or add a separate case, where we generate the kernel candidate from a config, i.e:

    ... 
    for i, cfg in enumerate(search_space):
        if not quiet and isatty:
            progress(0, i, total, len(errors))
        grid = grid_fn(cfg)
        kernel = kernel_fn(cfg)
        hints = hints_fn(cfg) if hints_fn is not None else {}
        updated_kernel = kernel.replace_hints(**hints)
        candidate = _TimingCandidate(
            config=cfg,
            grid=grid,
            kernel=updated_kernel,
            get_args=lambda _cfg=cfg: args_fn(_cfg),
        )
     ...

Describe any alternatives you have considered

No response

Additional context

No response

Contributing Guidelines

• I agree to follow cuTile Python's contributing guidelines
• I have searched the open feature requests and have found no duplicates for this feature request