TileCas

TileCas: Cascading Kernel Development Across Programming Abstractions

TileCas enables coding agents to develop accelerator kernels across editable programming abstractions. It connects high-level TileLang search with native Ascend C or HIP C++ refinement, using cross-level diagnosis, abstraction scheduling, and a heterogeneous search tree to retain and revisit candidate implementations.

Authors: Xinyu Xiao (Fudan University), Zihe Song (University of Texas at Dallas), Dezhi Ran (Peking University / Beijing Tongming Lake Information Technology Application Innovation Center), Wei Yang (Fudan University), and Tao Xie (Peking University / Fudan University).

Project website · arXiv · Results and materials

Source layout

Path	Purpose
kernelopt/protocol/	Shared optimization, diagnosis, export, evaluation, and candidate management
kernelopt/backends/ascend/	Ascend C / TileLang-Ascend adapters and evaluation tools
kernelopt/backends/rocm/	HIP / TileLang adapters and evaluation tools
kernelopt/backends/cuda/	Supporting CUDA export/profiling utilities; not a complete paper-platform launcher
kernelopt/agents/	Coding-agent CLI integration
skills/tilecas/	TileCas agent policy and iteration template
scripts/	Run staging, KDA integration, and foreground budget supervisor
configs/	Workload mappings, reference adapters, and routing configuration
native/	Native runtime support
thirdparty/	Selected KernelBench and AKO4ALL source, with their original licenses

Getting started

git clone https://github.com/TongmingLAIC/tilecas.git
cd tilecas
python3 scripts/run_experiment.py --help
python3 scripts/submit_kda_pair.py --help

The help commands require Python 3.10+ on Linux or macOS and do not launch experiments. Accelerator execution requires a Linux host with the appropriate SDK, PyTorch build, TileLang backend, and an authenticated coding-agent CLI. The framework uses source-tree entry points; it is not a standalone PyPI package.

Read installation and platform setup, then reproduction and entry points. KDA_REFERENCE.md documents the external KDA dependency and its dedicated launcher. Hardware SDKs, model services, and credentials are configured by the user.

How TileCas works

1. Develop a candidate at High (TileLang) or Low (native source).
2. Evaluate correctness and forward latency using the platform evaluator.
3. When the next change is unclear, inspect cross-level diagnosis with python scripts/diagnose.py in the generated experiment workspace.
4. Continue at the current level, lower a correct High candidate into editable native source, or restore a retained High candidate and explore another branch.

One agent conversation owns TileCas decisions across levels. Fixed Cascade is a separate comparison: one independent High agent, export and validation, then a fresh Low agent without High conversation history.

The internal mode name ours selects TileCas; existing kernelopt module names and command-line mode names are retained. Floating experimental inputs use N(0,1). Reference and candidate share the same inputs, seeds, and evaluation protocol. See protocol details.

Citation

See CITATION.cff, or use:

@misc{xiao2026tilecas,
  title = {TileCas: Cascading Kernel Development Across Programming Abstractions},
  author = {Xinyu Xiao and Zihe Song and Dezhi Ran and Wei Yang and Tao Xie},
  year = {2026},
  eprint = {2610.14740},
  archivePrefix = {arXiv},
  primaryClass = {cs.SE},
  url = {https://arxiv.org/abs/2610.14740}
}

License and contact

TileCas code is released under the MIT License. Bundled dependencies retain their own copyright notices and licenses; see THIRD_PARTY_NOTICES.md.

For project questions, open a GitHub issue. Correspondence: Wei Yang, yang_wei@fudan.edu.cn.