Vec-to-Cube-to-Vec Pattern

【免费下载链接】cannbot-skills CANNBot 是面向 CANN 开发的用于提升开发效率的系列智能体，本仓库为其提供可复用的 Skills 模块。 项目地址: https://gitcode.com/cann/cannbot-skills

Generic baseline only. For a2 (b3) kernels, prefer the a2-specific patterns under agent/references/patterns/ (e.g., a2-cube-vec.md, a2-cube-vec-cube.md) and read agent/references/constraints/a2-device.md for device-side rules.

Read this file when the kernel needs vec preprocessing, cube compute, and vec postprocessing in one fused flow.

Use this pattern when

• input data needs a vec-side transform first
• cube does the main matmul or cube-heavy stage
• vec performs the final output transform

Minimal flow

GM -> UB -> @vf -> UB -> L1 -> L0 -> L0C -> UB -> @vf -> GM

Ownership rule

This pattern has two cross-side handoffs:

• vec -> cube needs VcMutex
• cube -> vec needs CvMutex

Keep those two ownership edges conceptually separate. Do not let one counter or one fuzzy stage boundary blur them together.

What usually matters most

• separate stage counters
• clear stage ownership
• avoiding accidental reuse of one buffer lifetime across both handoff directions
• keeping validation incremental instead of building the whole fusion at once

Typical files to study

• agent/example/kernels/a5/vec_cube_vec_scale2_abs_add1_matmul.py

【免费下载链接】cannbot-skills CANNBot 是面向 CANN 开发的用于提升开发效率的系列智能体，本仓库为其提供可复用的 Skills 模块。 项目地址: https://gitcode.com/cann/cannbot-skills