Building-the-Regcmd-Graph: heterogeneous chaining (§4b) -- matmul->SDP via address aliasing; activation must be a separate task, not a fused matmul output; build on run_chain_i8; validated ork_mm_run_chain_i8_sdpsilu one-submit gate->silu-SDP
Add: Building the Regcmd Graph — segment/boundary methodology Replicable technique for turning any model + NPU into the fewest-submit on-NPU regcmd graph: segments (maximal contiguous NPU runs = one submit), the boundary taxonomy (hard/movable/free), the chaining mechanics (HW-chain within a segment, keepwarm one-mode, DMA-resident bsync-dedup, SW-chain precompiled across boundaries), and the 'collapse to #hard+1 mega-segments' principle. Qwen3-1.7B worked example: 6 segments -> 2, pivoting on softmax (the one load-bearing boundary = the 2b PPU RE). Portability notes (what varies per model/SoC vs the constant methodology).