| feat(pooling): add GlobalLpPool operator Co-authored-by: weixin_51153241<fanjie43@huawei.com> # message auto-generated for no-merge-commit merge: !6398 merge feature/add-global-lp-pool into master feat(pooling): add GlobalLpPool operator Created-by: weixin_51153241 Commit-by: weixin_51153241 Merged-by: cann-robot Description: ## 算子说明 Global LP Pool — 计算输入张量空间维度的 Lp 范数: y = (sum(|x|^p))^(1/p), keepdims=True ## 支持规格 - 输入: 4D (NCHW) 或 5D (NCD0D1D2) - DTypes: FP32, FP16, BF16 - 属性: p (float, 默认 2.0) - 架构: ascend950/arch35 - 接口: aclnn_exclude (直调) ## 文件清单 (20 files) - op_kernel/ — VF 级 kernel 实现 (双缓冲 + 二分归约树 + |x|^p + (1/p) root) - op_host/ — Tiling (ubFactor, 多核切分, cacheBuf 约束, empty tensor 分支) - op_graph/ — Proto 定义 - tests/ — Tiling 单测 (5 用例: 4D/5D/大通道/空张量) - README.md — 算子说明 See merge request: cann/ops-nn!6398 | 2 个月前 |