COMET

COMET is a framework for modeling and optimizing dataflow for compound operations on machine learning accelerators. COMET introduces a novel representation that explicitly models collective communication across spatial clusters, along with latency and energy cost models that account for GEMM and non-GEMM operation level dependencies within compound operations.

---

🛠️ Installation and Build

Prerequisites

Make sure the following packages are installed:

sudo apt install meson ninja-build
pip3 install meson

Setup Build Directory

From the repository root:

meson setup build --debug --warnlevel=2

Compile

cd build
meson compile

---

Usage

To run COMET, provide architecture, mapping, problem and config description files:

<path_to_comet>/build/comet --arch_file <arch_file> --mapping_file <mapping_file> --problem_file <prob_file> --constants_file <const_file>

Example

cd test/example_run
./run.sh

---

📂 Input Files

• arch_file — describes the target hardware architecture (e.g., compute array, memory hierarchy).
• mapping_file — defines how computations are mapped onto hardware resources.
• problem_file — specifies the workload characteristics (e.g., GEMM dimensions, tensor shapes).
• constants_file - specifies the value of constants used in mapping or problem files

---

📊 Output

COMET produces detailed logs and cost reports summarizing:

• Latency and throughput
• Energy and performance breakdowns

---

👥 Developers

• Shubham Negi — shubhamnegi1990@gmail.com
• Manik Singhal
• Aayush Ankit

---

📄 Citation

If you use COMET in your research, please cite our paper:

COMET: A Framework for Modeling Compound Operation Dataflows with Explicit Collectives Shubham Negi, Manik Singhal, Aayush Ankit, Sudeep Bhoja, and Kaushik Roy arXiv:2509.00599 (2025)

Click to expand BibTeX

@article{negi2025comet,
  title={COMET: A Framework for Modeling Compound Operation Dataflows with Explicit Collectives},
  author={Negi, Shubham and Singhal, Manik and Ankit, Aayush and Bhoja, Sudeep and Roy, Kaushik},
  journal={arXiv preprint arXiv:2509.00599},
  year={2025},
  url={[https://arxiv.org/abs/2509.00599](https://arxiv.org/abs/2509.00599)}
}