Sai Arthi R

• Leading team of 4 on customer project for front end graph compiler.
• Maintained graph compiler frameworks for major deep learning platforms, enabling inference.
• Upgraded requirements to latest PYPI version while addressing regressions.
• Optimized CI build and streamlined workflow processes.
• Ported computer vision kernels to designated hardware platform.
• Enhanced hardware utilization efficiency through kernel optimization.
• Resolved system pipeline bugs to improve overall stability.
• Leading another team of 4 in GenAI in networking research and creating POC for customer reach

Support and bug fix for Customer NN engine:

Project Goal: Fix bugs on CNN layer mismatch and write test suit to cover sanity and

regression tests for the layers and kernels .

Roles and Responsibility:

Understanding c and VecC kernels, load and store and basic instructions

• Fix bugs on Kernel mismatch and pipeline issues

• Worked on Onnx framework for reference pipeline

• Test suit development

• Testing failure cases and fixing it

Programming Languages: C, C++ IDE: Eclipse

Kernel Development for ONNX ops

Project Goal: The objective is to develop and test ONNX ops on AI accelerator

implemented in Halide language.

Roles and Responsibility:

• Implemented elementwise binary ops likr add, mul, sub in halide for 5D inputs

and tested the accuraccy with pytest

• Included quantization (scale per channel) for s8 dtype for the above ops.

• Implemented conv transpose pytest along with scale per channel for s8 dtype.

Programming Languages: Python, C++ IDE: Visual Studio

Porting kernel on DSP

Project Goal:

The objective is to build & optimize various CNN algorithms for a specific DSP with

real time performance. Roles and Responsibility:

• Implemented various kernels – 16bit Maxpool and gaussian kernels

• Optimizing the Kernels close to the theoretical estimate

• VLIW/SIMD based optimization

Programming Languages: C, C++. IDE: Visual Studio

Kernel development for Pytorch ops

Project Goal: The objective is to develop and test Pytorch 1.10 single ops on AI

accelerator.

Roles and Responsibility:

• Handling multiple dimensions for logsumexp

• Worked on subgraph generation and pass validation (force fallback and control

edge).

• Using Autocode for mm op

• Nearest1D forward and backward variants for upsample op

• Writing Gtest for single ops - Tanh, Hardsigmoid, Exp, Addcmul, CrossEntropy

and BinaryCrossEntropy with Logits.

Programming Languages: Python, C++ IDE: Visual Studio

• Created and implemented lesson plans that met the needs of the students curriculum.
• Handled Machine Learning and Data Science papers.