Gokulkrishna M

OnDevice ML @ CoreML Tools.

- Conducting weekly recitation classes on GPU Architecture and CUDA programming. - Grade assignments and projects for this course. - Worked under Professor Zahran.

• Developed prototype compiler using MLIR and Clang Front-end, simplifying AI Engine (NPU) programming by emitting optimized driver APIs. The tool captures AI Engine driver parameters and data structures using ClangAST and uses MLIR to emit code. • Fixed the Routing API for NPU Tiles by resolving errors, and bugs and writing unit tests for edge cases. This resulted in the Up-streaming of the API for the AI engine runtime driver code base for Versal Adaptive SoC.

• Programmed OpenCL GPU kernels for complex neural layers like group normalization and updated MLIR to support TfLite model conversion from Keras, enabling complete execution of Stable Diffusion models on Mobile GPUs. Engineered optimizations like convolution splitting, Quantization, GELU approximation, and fused softmax, which resulted in a 4.58x improvement in performance over CPU. This solution will serve as a base for infrastructure development for the acceleration of Gen-AI models on Samsung Mobile GPUs. • Designed an automatic recompilation & caching tool for OpenCL kernels on GPUs to reduce caching overhead and prevent crashes during GPU driver updates, which resulted in a 25% reduction in crash reports and a 23% increase in the productivity of engineers and the use case team. Received Spot award in Q3 2022 for developing this tool. • Accelerated 10 use cases on Samsung flagships (S23 devices) to improve user experience, achieved a 20% improvement in load time and 10% in execution time and overall performance improvement of 1.4x (S23) and 1.3x (S24).

• Engineered a Profiling tool for calculating the layer-wise performance of kernels in ArmNN during inference of ML model. Resulting in the faster diagnosis of performance degradation and improved the productivity of the team by 30%. • Collaborated with Galaxy RAW team to solve greenish tinge issues on RAW images by implementing new normalization methods to support float16 quantized precision execution. This improved inference time by 34% over the previous method. • Directly helped in enabling over 15 USP camera and gallery features deployed on Samsung flagship S22 and S23 devices. Contributing to over 1.5x optimizations for speed, memory and battery.

Evaluated the improvement in emoji prediction by incorporating keystroke statistics in NLP sentiment analysis algorithm, resulting in 10% improvement in accuracy on personalized dataset.

Engineered Android Application, which uses firebase cloud tools to create real time service for engineers to handle service requests, resulting in improved and standardized system for engineers.