arXiv subjects

Sanjay Motilal Ladwa

Publications and source records attributed to Sanjay Motilal Ladwa.

3 recordsLinked to original sources

arXiv2026

Hexagon-MLIR: An AI Compilation Stack For Qualcomm's Neural Processing Units (NPUs)

In this paper, we present Hexagon-MLIR,an open-source compilation stack that targets Qualcomm Hexagon Neural Processing Unit (NPU) and provides unified support for lowering Triton kernels and PyTorch models . Built using the MLIR framework, our compiler applies a structured sequence of passes to exploit NPU architectural features to accelerate AI workloads. It enables faster deployment of new Triton kernels (hand-written or subgraphs from PyTorch 2.0), for our target by providing automated compilation from kernel to binary. By ingesting Triton kernels, we generate mega-kernels that maximize data locality in the NPU's Tightly Coupled Memory (TCM), reducing the bandwidth bottlenecks inherent in library-based approaches. This initiative complements our commercial toolchains by providing developers with an open-source MLIR-based compilation stack that gives them a path to advance AI compilation capabilities through a more flexible approach. Hexagon-MLIR is a work-in-progress, and we are continuing to add many more optimizations and capabilities in this effort.

cs.PL↗

arXiv2014

SmartConnect: A System for the Design and Deployment of Wireless Sensor Networks

We have developed SmartConnect, a tool that addresses the growing need for the design and deployment of multihop wireless relay networks for connecting sensors to a control center. Given the locations of the sensors, the traffic that each sensor generates, the quality of service (QoS) requirements, and the potential locations at which relays can be placed, SmartConnect helps design and deploy a low- cost wireless multihop relay network. SmartConnect adopts a field interactive, iterative approach, with model based network design, field evaluation and relay augmentation per- formed iteratively until the desired QoS is met. The design process is based on approximate combinatorial optimization algorithms. In the paper, we provide the design choices made in SmartConnect and describe the experimental work that led to these choices. We provide results from some experimental deployments. Finally, we conduct an experimental study of the robustness of the network design over long time periods (as channel conditions slowly change), in terms of the relay augmentation and route adaptation required.

cs.NI↗

arXiv2014

A Fast and Accurate Performance Analysis of Beaconless IEEE 802.15.4 Multi-Hop Networks

We develop an approximate analytical technique for evaluating the performance of multi-hop networks based on beaconless IEEE 802.15.4, a popular standard for wireless sensor networks. The network comprises sensor nodes, which generate measurement packets, relay nodes which only forward packets, and a data sink (base station). We consider a detailed stochastic process at each node, and analyse this process taking into account the interaction with neighboring nodes via certain time averaged unknown variables (e.g., channel sensing rates, collision probabilities, etc.). By coupling the analyses at various nodes, we obtain fixed point equations that can be solved numerically to obtain the unknown variables, thereby yielding approximations of time average performance measures, such as packet discard probabilities and average queueing delays. The model incorporates packet generation at the sensor nodes and queues at the sensor nodes and relay nodes. We demonstrate the accuracy of our model by an extensive comparison with simulations. As an additional assessment of the accuracy of the model, we utilize it in an algorithm for sensor network design with quality-of-service (QoS) objectives, and show that designs obtained using our model actually satisfy the QoS constraints (as validated by simulating the networks), and the predictions are accurate to well within 10% as compared to the simulation results.

cs.NI↗