Optimizes MNN Vulkan backend kernels and integrates new Vulkan features on mobile GPUs.
npx skills add https://github.com/alibaba/MNN --skill vulkan-optimize
Please help me install this Agent Skill. Skill: vulkan-optimize Source: https://github.com/alibaba/MNN/tree/main/skills/vulkan-optimize Command: npx skills add https://github.com/alibaba/MNN --skill vulkan-optimize
Optimizing GPU compute shaders in the MNN Vulkan backend is error-prone: edited GLSL .comp files silently do nothing without regenerating SPIR-V arrays, stale pipeline caches cause segfaults, and prefill is often CPU-scheduling-bound rather than GPU-bound, so naive kernel tuning yields no end-to-end gain. This Skill provides a disciplined benchmark-optimize-integrate workflow with correctness oracles and interleaved A/B measurement on real Android devices.
Ask the AI to profile and optimize the Vulkan attention or convolution kernels of an MNN LLM on your connected Android device, following the benchmark, kernel optimization, and integration steps.
High-intent search queries and answers about installing and using this skill.
Profile with MNN_GPU_TIME_PROFILE=ON to get op-level and shader-level GPU timings, then determine whether the bottleneck is CPU command recording or GPU kernels. Apply matched techniques such as cooperative matrix, subgroup reductions, or epilogue fusion, verifying each with interleaved A/B benchmarks.
MNN does not compile .comp files at build time; it loads SPIR-V byte arrays embedded in AllShader.cpp by makeshader. You must regenerate the arrays for changed shaders, and confirm you edited the correct buffer or image backend tree selected by MNN_VULKAN_IMAGE.
The persisted VkPipelineCache in tmp/mnn_cachefile.bin becomes stale when shaders change, causing a segfault on load. Delete the cache file with rm tmp/mnn_cachefile.bin before every test run after modifying shaders.
No. Cooperative matrix in MNN is gated to Adreno GPUs reporting supportCoopMat, and it only has hardware definitions for int8 and int4 weights. Unsupported devices must fall back to the original shader path via runtime feature detection.
MNN Vulkan prefill is often CPU-scheduling-bound, so GPU kernel speedups show no end-to-end gain. Compare cumulative GPU kernel time against wall time first; if CPU-bound, use indirect batch recording or operator fusion instead of kernel tuning.
Use interleaved A/B testing where baseline and optimized builds alternate back-to-back, because device thermal drift causes 8-10% variance across sessions. Warm up first, clear the pipeline cache each round, and measure clean builds without the GPU profiler enabled.
asnar00Build and deploy Android APKs to USB-connected devices via ADB.
tile-aiValidate Python and C++ code formatting against CI rules and generate JSON/Markdown reports.
K-Dense-AIPorts CPU-bound scientific Python workloads to NVIDIA GPU libraries and validates correctness and speedup.
VMAFxScaffold a GPU backend package for libvmaf with build wiring and smoke tests.
