Developing a deep learning compiler stack that takes neural network descriptions (CNNs/RNNs) created in frameworks such as Caffe, PyTorch, TensorFlow, etc. and converts them into code suitable for execution on special-purpose and embedded platforms.

Use modern compiler frameworks such as LLVM and MLIR.

Developing optimized implementations of a variety of neural-network operations and integrating them into a runtime framework

Developing new optimization techniques and algorithms to efficiently map CNNs onto a wide range of Xtensa processors and specialized hardware.

Benchmarking end-to-end network performance on a variety of DSP and special-purpose accelerator platforms.

Enhancing the framework to improve overall functionality and performance on the various hardware platforms.

Devising multiprocessor/multicore partitioning and scheduling strategies.