Reference model
The executable definition of arithmetic, layouts, activations, and edge cases.
HASLAB is building an open inference engine from exact software behavior outward. Run the numerical model, command simulator, and conformance tests first. Then explore the calibrated YOLOv8n workload and its path through the learned detection head and declared host tail.
The executable definition of arithmetic, layouts, activations, and edge cases.
The executable definition of v0 commands, memory behavior, completion, reset, and errors.
46 independent binary fixtures another implementation can consume without private model helpers.
With Git, Python 3, and Make installed, clone the repository and run its checks. The source-independent suite covers reference, simulator, compiler, runtime, and conformance behavior without downloading model weights or COCO images.
This exercises the 46 stored binary fixture cases and the simulator runner.
Passing demonstrates agreement for tested software cases. It is not an FPGA result, silicon result, or performance measurement.
The pinned YOLOv8n 320×320 workload has FLOAT and calibrated INT8 software baselines. Nodes 0–217 compile to 412,092 commands; the synthetic diagnostic run matches all 6,931,200 stored values across 106 boundaries.
The INT8 software proxy loses 0.887 mAP points, within the one-point budget. Separately, the six-input command-path check matches 37,900,800 stored integer values exactly.
Understand the results →Use the workload guide to prepare the pinned Python 3.11 export environment, model, and calibration artifacts. Then run this from the repository root:
PYTHONPATH=reference:simulation:compiler:runtime \
python benchmarks/tools/compile_yolov8n_detection_head.pyArtifact setup + reproduction guide ↗Each path is useful on its own. Start small, follow the linked source, and return to the architecture page when you want the full-system view.
When you are ready for DMA, SRAM, tensor operations, the vector path, planned RISC-V control, and the route toward FPGA, continue to the architecture guide.
Architecture guide →The stable v0 ABI still awaits independent review. Useful hardware work starts with reviewed contracts, independent vectors, and a narrow vertical slice—not a speculative full accelerator.