Pin the input
The workload guide records model hash, export environment, graph audit, and calibration artifacts.
The compiler validates accelerator nodes 0–217 of the pinned opset-13 graph: fused Conv → Sigmoid → Mul patterns, 1×1 and 3×3 convolutions, split views, scaled residual and non-residual C2f blocks, MaxPool, exact 2× upsampling, and channel concatenation. It extracts constants, uses recorded activation scales and SiLU lookup tables, packs INT8 weights, derives INT32 bias records, and writes deterministic experimental .hxb packages. Its reusable scheduler chooses SRAM-fit tiles, streams wide weights and concat parameters, preserves long-lived skip tensors, and applies deterministic lifetime allocation.
The workload guide records model hash, export environment, graph audit, and calibration artifacts.
The compiler guide explains supported patterns and package generation; the format note details the graph schedule.
The Python runtime validates hashes and relocations, binds an exact-size input, submits commands, exposes diagnostic or release outputs, and returns completion or a structured fault.
One graph package submits 412,092 candidate ABI commands through the functional simulator. Its 44,598,232 scheduled DMA bytes and FIFO refill counts describe software command behavior, not cycles, latency, or FPGA throughput. Diagnostic mode retains comparison boundaries; release mode uses lifetime reuse and exposes three raw INT32 learned-head tensors. Practical command delivery remains a pre-RTL design question.
DMA and boundary fills bring input patches and constants into modeled local storage; wide weights and concat parameters stream when their full records exceed local SRAM.
CONV_I8 and EPILOGUE handle quantized convolutions and SiLU; exact MAXPOOL5_I8 and UPSAMPLE2_I8 cover SPPF and both upsampling stages.
The synthetic diagnostic run matches 6,931,200 stored values across 106 boundaries; release mode matches all 302,400 final INT32 values. Six further inputs match 37,900,800 stored values exactly. Inspect the report and scope.
The package schema and Python runtime are experimental research interfaces.
The development plan calls for a faster command-equivalent runner, verified against the conformance corpus and six diagnostic inputs, before a 5,000-image COCO command-path evaluation. The 412,092-command complete schedule also needs a practical delivery decision before interface RTL.