[FEA]: CUDA C or PTX Injection

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

New Feature

How would you describe the priority of this feature request?

Low (would be nice)

Please provide a clear description of problem this feature solves

According to the test_bytecode.py file, cuTile supports launching kernels provided in CUBIN format, which enables execution of CUDA C kernels compiled offline. However, this creates a strict separation between Python-authored cuTile kernels and CUDA C kernels. Users must choose one approach or the other, with no supported mechanism to combine them. As a result, it is difficult to reuse existing CUDA C or PTX code, or to optimize performance-critical regions within an otherwise Python-based cuTile kernel.

Feature Description

Add support for embedding or injecting CUDA C or PTX code into a Python-authored cuTile kernel. This would enable a hybrid programming model where most kernel logic is expressed in Python, while selected sections can be implemented in CUDA C or PTX for fine-grained performance tuning or access to low-level hardware features. This capability would improve cuTile's flexibility, allow reuse of existing CUDA C/PTX code, and make cuTile a more powerful tool for advanced CUDA kernel development.

Describe your ideal solution

Provide an API that allows directly inserting CUDA C or PTX instructions into a Python-authored cuTile kernel, analogous to asm volatile(...) in CUDA C. This API would act as a low-level escape hatch, enabling users to inline raw code at specific points in the kernel.

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

Inizia da test/test_bytecode.py e dalla copertura esistente dell’avvio dei kernel CUBIN per capire come vengono attualmente supportati i kernel compilati esternamente. Definisci con i maintainer il confine previsto per l’iniezione di CUDA C/PTX e il comportamento dell’API, quindi aggiungi test che mostrino che il codice incorporato può coesistere con kernel cuTile scritti in Python.