Thanks for sending an requirement! Please fill in the following template to help quickly solve your problem.
depthwiseConv2d在Conv2d通用kernel模板下scalar问题很大，L1 L0层级循环流水被打断，对比竞品性能不达标
Conv正向
qwen3-next中decode中有相当多的depthwise
通过提前确定计算循环轴的数量单独实现depthwiseconv2d的kernel
Thanks for sending an requirement! Please fill in the following template to help quickly solve your problem.
Backgroud（背景信息）
depthwiseConv2d在Conv2d通用kernel模板下scalar问题很大，L1 L0层级循环流水被打断，对比竞品性能不达标
Origin（信息来源）
Conv正向
Benefit / Necessity （价值/作用）
qwen3-next中decode中有相当多的depthwise
Design（设计方案）
通过提前确定计算循环轴的数量单独实现depthwiseconv2d的kernel