MiniAIC

Build a AI compiler from scratch. It only does graph optimization, no kernel generation: takes ONNX, writes a flatbuffer (.gynn) for MiniNN to run.

Build

cmake -B build -S .
cmake --build build

Usage

# optimize and export: models/xxx.onnx -> models/xxx.gynn (plus a .json for Netron)
./build/miniaic/frontend/main models/mobilenetv2-10.onnx

# skip the optimizations (keeps only the required shape inference) to see what they do
./build/miniaic/frontend/main models/mobilenetv2-10.onnx --no-opt

# run it with MiniNN (build MiniNN's lib first, see demo/CMakeLists.txt for the paths)
./build/demo/demo models/mobilenetv2-10.gynn

Optimization passes (miniaic/opt/)

Run in order, each one depends on the result of the previous:

pass	what it does
shape_inference	infers the shape of every tensor, and evaluates the constants on shape chains
constant_folding	evaluates and removes nodes whose result is already known at compile time
fuse_conv_clip	fuses Conv + Clip into a single Conv (the clip bounds become attributes)
dce	removes unreachable nodes, reclaims dead tensors and remaps tensor indices

Before serializing, the graph is validated (unknown shapes, tensor ownership, graph outputs must have a producer), and the .gynn is not written if the check fails.

Layout

directory	contents
miniaic/frontend/	read ONNX, build the IR, run the passes, export
miniaic/ir/	internal IR: graph / node / tensor / type
miniaic/opt/	graph optimization passes
miniaic/serializer/	IR -> .gynn, Netron JSON
proto/	how to generate the ONNX protobuf bindings
fbs_schema/	how to generate the .gynn flatbuffers schema
demo/	run inference with MiniNN

More background: docs/README.md.