Digant Desai
Software Engineer at Meta
Austin, Texas, United States
Summary
🤩
Rockstar🎓
Top SchoolDigant Desai is a software engineer with 11 years of experience specializing in low-level programming, OS internals, and CPU performance modeling, currently contributing to PyTorch Edge at Meta from Austin. He has a strong track record in performance engineering across industry leaders—optimizing neural network inference in XNNPACK for ARM aarch64 and enhancing PyTorch quantization and profiler support. Prior roles at Arm and Qualcomm focused on microarchitecture exploration and server CPU optimization, linking software stacks to hardware behavior. His research background produced energy-aware scheduling and CGRA accelerator designs, reflecting a rare blend of academic rigor and production-grade systems engineering. Notably, his open-source work includes assembly-level SIMD optimizations that materially speed up QU8 GEMM kernels on Cortex-A55 processors.
11 years of coding experience
11 years of employment as a software developer
M.S., Elecrical Engineering, M.S., Elecrical Engineering at Arizona State UniversityB.Tech, Electronics, B.Tech, Electronics at NIRMA UNIVERSITY
English, Hindi, Gujarati
Github Skills (24)
pytorch10
performance-monitor10
performance-analytics10
python10
matrix-multiplication10
cpu10
performance-measurement10
xpack10
performance-analysis10
optimisation10
neural-network10
performance-tuning10
simd10
quantization10
performance-monitoring10
Programming languages (7)
Github contributions (5)
pytorch/pytorch
Oct 2021 - Dec 2022
Tensors and Dynamic neural networks in Python with strong GPU acceleration
Role in this project:
Backend Engineer
Backend EngineerContributions:138 reviews, 167 commits, 102 PRs in 1 year 2 months
Contributions summary:Digant primarily contributed to the PyTorch library, specifically focusing on quantization and performance optimization. Their work included implementing clamping for quantized weights, fixing per-channel weight observers, and adding support for performance event names in the profiler. Furthermore, the user added Linux Perf support to the profiler and enhanced the XNNPACK library by enabling various kernels to improve performance. They also added tests related to the changes.
pythongpu-accelerationdeep-learninggpunumpy
google/XNNPACK
Nov 2021 - Feb 2022
High-efficiency floating-point neural network inference operators for mobile, server, and Web
Role in this project:Performance Engineer
Performance EngineerContributions:59 reviews, 10 commits, 21 PRs in 2 months
Contributions summary:Digant focused on optimizing the performance of neural network inference operators within the xnnpack library. Their commits primarily involved enabling and refining FP32 requantization variants for QU8 GEMM kernels, specifically for aarch64 NEON dot product implementations on various ARM Cortex-A55 processors. These changes involved modifications to assembly code, specifically focusing on the efficient use of SIMD instructions to improve the speed of matrix multiplications, demonstrating a deep understanding of low-level optimizations.
multithreadingsimdtensorflowcpufloating
Find and Hire Top DevelopersWe’ve analyzed the programming source code of over 60 million software developers on GitHub and scored them by 50,000 skills. Sign-up on Prog,AI to search for software developers.