Extensions to Torch-MLIR

This repository contains Xilinx extensions to the torch-mlir ATen dialect to enable expressing the input and output of partial results. Using these extensions, a prototype streaming dataflow exploration tool scans the design space for compute/memory/computation balanced partitioning of CNNs on the AIE array of Xilinx’ Versal devices.

Getting Started

Github sources

Generated code documentation

XTen Dialect
XTen Passes