Optimizes and validates MNN Hexagon DSP backend kernels with device-measured profiling.
npx skills add https://github.com/alibaba/MNN --skill hexagon-optimization
Please help me install this Agent Skill. Skill: hexagon-optimization Source: https://github.com/alibaba/MNN/tree/main/skills/hexagon Command: npx skills add https://github.com/alibaba/MNN --skill hexagon-optimization
Optimizing MNN's Hexagon/HVX/HMX DSP backend is error-prone: intuition-driven micro-optimizations often fail, host/DSP build mismatches silently corrupt measurements, and v79/v81 architecture differences invalidate conclusions. This Skill encodes hard-won measurement discipline, phase-level profiling, and verified bottleneck patterns so changes are driven by real device data.
Use the hexagon-optimization skill to analyze the DSPOpType profile output of my MNN model and identify which phase of the prefill GEMM operator to optimize.
High-intent search queries and answers about installing and using this skill.
Enable the gated phase profilers (HTP_MM_PHASE_PROFILE, HTP_QATTN_PHASE_PROFILE, HTP_WATTN_PHASE_PROFILE) in both paired files, rebuild, and read the pseudo-operator entries in the DSPOpType table. Keep profilers off when reporting throughput numbers.
Measure first with phase decomposition, then target the dominant phase: merge small DMA transfers into fewer larger ones, fix issue-bound writebacks with more parallel issue, and avoid partial-slot writes that degrade to read-modify-write. Validate on both v79 and v81 architectures.
Real devices have one-sided noise from background system services, with prefill around ±4% and decode ±7%. Use best-of-N or median instead of mean, re-test baselines at the start and end of a session, and always rebuild host code for every measurement point.
Yes. Per-operator v79/v81 time ratios cluster around 1.1-1.3x, HMX locking differs (shared on v79, exclusive on v81), and weight layouts are architecture-specific. Task partition counts must not collide with thread counts or wave quantization can erase algorithmic gains.
cDSP crashes can come from VTCM over-allocation (vtcm_seq_alloc has no bounds checking), stale cache lines overwriting host uploads, or 4096-token prefills that hard-reboot the device. Check tombstones and ramdumps under /data/tombstones and /data/vendor/ before concluding anything.
Do not use CPU fallback or operator fusion as an optimization, do not gate code paths on environment variables, and do not retry disproven directions like dmlink chained DMA, non-temporal stores in prefill GEMM, or weight staging double-buffering, which previously caused device reboots.
Arm-ExamplesCross-compile LLM libraries for Android AArch64 using CMake presets and the Android NDK.
alibabaOptimize MNN operators for ARM CPUs with multi-threading and assembly kernels.
ruvnetImports Claude Code auto-memory files into AgentDB with ONNX embeddings for unified semantic search.
choice5346Run LLM inference on CPUs, Apple Silicon, and non-NVIDIA GPUs with llama.cpp.
