- This repository contains the hardware implementation of a streamlined LSTM layer for use with FINN, leveraging
finn-hlslibbuilding blocks. - Paper link: https://arxiv.org/pdf/2506.20810
- The top-level computation is defined in
qlstm_top.cpp, which represents the operations derived from the streamlined LSTM QONNX graph. - Layer-specific parameters (size of inputs, outputs, lookbacks and datatypes) are provided in
pipeline-lstm-header.h.
The generated hardware can be exported as a synthesizable IP block, which integrates seamlessly with other FINN-compatible layers such as convolutional and fully connected layers. This enables the extension of FINN to support Recurrent Neural Networks (RNNs), particularly LSTM-based architectures.
- Streamlined and synthesizable LSTM design
- AXI-Stream interface for easy integration
- Compatible with FINN-generated QONNX models
- Modular IP for combining with other layers