iiinova — where your digital self keeps evolving

Kabilan Mahathevan @kabilan-m.bsky.social · 27/09/2026

Splyce transforms this unpredictable control flow into a predictable dataflow, eliminating branch mispredictions while improving instruction-level parallelism. This work has been accepted to #PACT26, and I’ll be presenting it on October 21st! 📄 Preprint: doi.org/10.48550/arX...

doi.org

Splyce: SIMD Vectorization of Sparse Coiteration

Sparse tensor contractions are bottlenecked by sparse-sparse coiteration loops that resist standard loop vectorization. We present Splyce, an auto-vectorization framework in MLIR that overcomes this t...

000

Kabilan Mahathevan @kabilan-m.bsky.social · 27/09/2026

Unfortunately, data-dependent branches and bounds make this code pattern extremely difficult for traditional auto-vectorization. #Splyce is an auto-vectorization framework, implemented as an optimization pass that intercepts the MLIR Sparsifier pipeline.

100

Kabilan Mahathevan @kabilan-m.bsky.social · 27/09/2026

Sparse tensor contractions are fundamental to many HPC, graph analytics, and AI/ML workloads. When both input tensors are sparse along the contracting dimension, the generated code requires Sparse-Sparse coiteration over the shared index.

100

Kabilan Mahathevan @kabilan-m.bsky.social · 25/02/2026

I’ll be presenting the initial design and preliminary results at the 2026 Fuzzing Workshop ( #FUZZING26 ) in San Diego, co-located with #NDSS26. If you're building a sparse tensor compiler and want to integrate or stress-test it, come find me at #FUZZING26 — I would be happy to talk.

github.com

GitHub - KabilanMA/TenSure: Modular Black-box fuzzing framework for STCs

Modular Black-box fuzzing framework for STCs. Contribute to KabilanMA/TenSure development by creating an account on GitHub.

100

Kabilan Mahathevan @kabilan-m.bsky.social · 25/02/2026

TenSure uses metamorphic testing across multiple data-layout formats to systematically stress-test sparse tensor compiler implementations. It takes Einsum notation as a frontend abstraction, making it straightforward to integrate with existing systems.

github.com

GitHub - KabilanMA/TenSure: Modular Black-box fuzzing framework for STCs

Modular Black-box fuzzing framework for STCs. Contribute to KabilanMA/TenSure development by creating an account on GitHub.

200

Kabilan Mahathevan @kabilan-m.bsky.social · 25/02/2026

We’re introducing #TenSure (github.com/KabilanMA/Te...), the first black-box fuzzer built specifically for sparse tensor compilers. This work is a collaboration with Yining Zhang, Muhammad Ali Gulzar, and @kirshanthans.bsky.social at Virginia Tech.

github.com

GitHub - KabilanMA/TenSure: Modular Black-box fuzzing framework for STCs

Modular Black-box fuzzing framework for STCs. Contribute to KabilanMA/TenSure development by creating an account on GitHub.

221

Kabilan Mahathevan @kabilan-m.bsky.social · 08/12/2024

I presented my first conference paper at the 31st Asia-Pacific Software Engineering Conference (APSEC-2024) in Chongqing, China. It was an exciting and memorable journey!

050