Add new operators to MNN through a staged workflow covering data definitions, output shapes, computation, and tests. It prioritizes combining existing operators, with CPU implementations and additional hardware backends used when needed.
为 MNN 框架新增算子。包含 Schema 定义、形状计算、几何计算、后端实现、单元测试的完整 TDD 流程。分 5 步执行，每步有独立测试标准。
Review the instructions the agent will receive before you install.
触发条件：当用户请求为 MNN 添加/实现一个新的算子时触发。常见表述包括："添加xxx算子"、"实现xxx op"、"支持xxx操作"等。
本 SKILL 指导 AI Agent 为 MNN 框架新增算子支持。采用 TDD（测试驱动）模式，每一步都有明确的测试标准。
优先级顺序：几何计算 > 后端实现
严禁访问以下目录：
schema/private/和source/internal/，包含内部私有代码，不得读取、修改或引用。
名字必须等于真实物理量：新增或修改任何后端算子 / kernel 代码时，变量、字段、宏名指代的必须是代码里真正装的那个量——
blockSize是"块内元素数"还是"每行块数"、step是元素数还是字节数、shape 门禁的自变量是通道数还是 tile 数。名字骗人不会报错（编译、对拍、单测都过），但后续基于它写的阈值、门控、分档会一起错。写完对着实际索引 / 指针算术复核一遍，不符就改名。
| 目录/文件 | 作用 | 何时修改 |
|---|---|---|
schema/default/MNN.fbs | 算子类型和参数定义（开源版） | 每个新算子都需要；内部仓库实际生成走 schema/private/，计划开源发布的算子两份都要改，详见 step1-schema.md §1.5 |
schema/default/CaffeOps.fbs | Caffe 框架算子参数 | 有参数时 |
schema/default/TensorflowOp.fbs | TF 框架算子参数 | 有参数时 |
source/shape/Shape*.cpp | 形状（维度）计算 | 输出形状与输入不同时 |
source/geometry/Geometry*.cpp | 几何计算（算子拆解） | 优先实现 |
source/backend/cpu/CPU*.cpp/.hpp | CPU 后端实现 | 无法几何拆解时 |
source/backend/metal/Metal*.cpp/.hpp/.metal | Metal 后端实现 | 扩展到 Metal |
source/backend/opencl/execution/*.cpp/.h/.cl | OpenCL 后端实现 | 扩展到 OpenCL |
source/backend/vulkan/image/execution/Vulkan*.cpp/.hpp | Vulkan 后端实现 | 扩展到 Vulkan |
source/backend/cuda/execution/*.cu/.cuh | CUDA 后端实现 | 扩展到 CUDA |
include/MNN/expr/ExprCreator.hpp | 表达式 API 函数 | 需要用户可调用时 |
test/op/*Test.cpp | 单元测试 | 每个新算子都需要 |
tools/script/register.py | 注册脚本 | 添加 shape/geometry/backend 后运行 |
┌──────────────────────────────────────────────────────────┐
│ 步骤 1: Schema 定义 (step1-schema.md) │
│ 输入: 算子名称、语义、输入输出定义 │
│ 输出: MNN.fbs 中的 OpType + 可选的参数 table │
│ 测试: flatc 编译通过 │
├──────────────────────────────────────────────────────────┤
│ 步骤 2: 形状计算 (step2-shape.md) │
│ 输入: 算子的输入输出维度关系 │
│ 输出: source/shape/Shape*.cpp │
│ 测试: register.py 运行成功 + cmake 编译通过 │
├──────────────────────────────────────────────────────────┤
│ 步骤 3: 几何计算或 CPU 实现 (step3-compute.md) │
│ 输入: 算子的计算逻辑 │
│ 输出: Geometry*.cpp 或 CPU*.cpp │
│ 测试: register.py + 编译通过 │
├──────────────────────────────────────────────────────────┤
│ 步骤 4: 单元测试 (step4-test.md) │
│ 输入: 算子的预期行为 │
│ 输出: test/op/*Test.cpp │
│ 测试: ./run_test.out op/MyOp 全部通过 │
├──────────────────────────────────────────────────────────┤
│ 步骤 5: 扩展后端与优化 (step5-optimize.md) │
│ 输入: 步骤 4 通过 │
│ 输出: Metal/OpenCL/Vulkan/CUDA 实现 │
│ 测试: 各后端单元测试通过 │
└──────────────────────────────────────────────────────────┘
| 情况 | 需要执行的步骤 |
|---|---|
| 输出形状与输入完全一致，可几何拆解 | 1 → 3(几何) → 4 |
| 输出形状与输入不同，可几何拆解 | 1 → 2 → 3(几何) → 4 |
| 输出形状与输入不同，需要后端实现 | 1 → 2 → 3(CPU) → 4 → 5 |
| 输出形状与输入完全一致，需要后端实现 | 1 → 3(CPU) → 4 → 5 |
现在请打开 skills/add-new-op/step1-schema.md，开始步骤 1。
No reviews yet. Tell others what you thought.
No login needed. Share how it worked for you, ask a question, or note where it did not run.
No comments yet. Be the first to write one.
Triage incorrect MNN outputs and regressions by matching symptoms to focused debugging guides. Use backend comparisons, minimal reproductions, and controlled experiments to investigate memory issues, quantization, concurrency, numeric limits, caches, and inconsistent results across runs.
Find the right workflow for MNN CPU work, from investigating slow execution to implementing calculation routines and planning tests. Compare ARM, x86_64, and RISC-V backend structures before choosing the relevant guidance.
Develop or port CPU kernels, the low-level calculation routines used by MNN. Follow staged implementation guidance, check data layout agreements, register functions, and verify correctness across processor instruction sets and numerical precision settings.
Find why MNN CPU inference is slow by examining threads, scheduling, memory layout, and the processor instruction path actually used. Operator-level measurements and comparisons with other inference frameworks help identify which layer needs improvement across ARM, x86_64, and RISC-V.
Optimize and troubleshoot MNN's Hexagon DSP backend, which runs computation on a dedicated processor. The workflow uses device measurements to identify bottlenecks, compare v79 and v81 architectures, verify correctness, and investigate crashes or memory usage.
Guide MNN Metal backend development and optimization, from finding performance bottlenecks to changing GPU computation, combining operators, and adjusting execution scheduling. A task-based document index directs work toward relevant implementation rules, benchmarks, and correctness checks.