Alkaid: Distributed Arithmetic for Machine Learning

alkaid is a lightweight compiler for generating low-latency static-dataflow kernels for FPGAs. It traces quantized arithmetic into ALIR, applies distributed-arithmetic optimization through CMVM where useful, and emits RTL or HLS projects.

The supported code generation targets are:

• RTL: Verilog and VHDL 2008, with optional pipelining.
• HLS: Vitis HLS, HLSlib, and oneAPI-style C++.
• XLS: experimental Verilog generation and JIT execution through xls-python.

Models can be described with the NumPy-like symbolic tracing API, loaded from serialized ALIR JSON/JSON.GZ, or converted through plugins. The package includes top-level keras, torch, and alkaid example tracer plugins; Keras and Torch support can also be extended with second-level alkaid_keras and alkaid_torch entry points.

Installation

pip install alkaid

Binary wheels are published for Linux x86_64 and macOS ARM64. Building from source requires Python 3.10 or newer, NumPy, meson-python, and a C++20 compiler with OpenMP support.

pip install --no-build-isolation -e '.[tests]'

Use pip install 'alkaid[docs]' for documentation dependencies. Install xls-python when using the XLS backend outside the test extra.

CLI

alkaid convert model.keras /tmp/rtl_output --flavor verilog --latency-cutoff 5
alkaid convert model.json.gz /tmp/hls_output --flavor vitis
alkaid report /tmp/rtl_output --sort-by comb_metric

See the documentation for tracing, plugin, ALIR, and backend details.

License

LGPLv3. See the LICENSE file for details.