In this talk, we will give a walk-through of our MLIR-based graph compiler optimizing TensorFlow Lite models to be deployed on the xcore.ai microcontroller. We focus specifically on MLIR passes for memory usage reduction, such as offline memory planning, operator splitting, and streaming constants from flash, along with other passes. We leverage open-source projects such as MLIR, Tensorflow, and tflite-micro-compiler to create a small executable within various resource constraints. In contrast to other compilers in the LLVM world, we do not lower to LLVM IR. Instead, we produce C++ source code using tflite-micro-compiler, which is compiled and executed by our toolchain. We will go through some of the challenges in our journey and future plans. All code for the graph compiler and runtime is available publically on GitHub.

Max Bartel