变更摘要

本 PR 完整回退 [GE #5200] 合入 develop 的提交 5dfaef763，恢复 GE 旧有 ordinary AICPU 常量折叠实现、target、stub、UT/ST 与交付项，并恢复原有 HostCpuEngine 加载顺序。回退原因在于迁移后 bottom ordinary 路由构造 AICPU NodeDef 时改用实际常量 Tensor 提供 input shape/format/dtype，导致与节点输出描述冲突、AICPU 失败后 fallback 到 GE built-in，行为与迁移前不一致。核心动作是删除迁移期引入的 libopconstant_folding.so 装载链路，改回 libaicpu_const_folding.so 及 libops_host_cpu.so 的延迟初始化与兜底逻辑，并重新引入 aicpu_const_folding 目录下的 folding、host_cpu_kernel_wrapper 等实现及配套构建、打包、测试项。

主要改动

• 恢复 HostCpuEngine 加载顺序与库名: base/host_cpu_engine/host_cpu_engine.cc/.h 将库名改回 kAicpuConstFoldingName = "libaicpu_const_folding.so"，删除 LoadOpConstantFoldingLib，改为先 dlopen 但延迟调用 libops_host_cpu.so 的 Initialize，随后加载 libaicpu_const_folding.so 并执行 Initialize，失败时回退调用 InvokeLibInitialize 兜底。
• 简化 DlopenLib 接口: 去掉 DlopenLib 的 bool global 参数，固定使用 MMPA_RTLD_NOW | MMPA_RTLD_GLOBAL 打开标志，host_cpu_engine.h 同步更新声明。
• 新增 ordinary AICPU 常量折叠实现: 新增 compiler/engines/cpu_engine/cpu_engine/aicpu_const_folding/folding.cc/folding.h，导出 InitCpuConstantFoldingNew 与 CpuConstantFoldingComputeNew，实现扫描并加载 libopconstant_folding_*.so、V1/V2 算子枚举与 op_type 反向索引构建、NodeDef 输入输出与属性转换及 V1/V2 计算路由。
• 新增 Host CPU kernel 包装层: 新增 host_cpu_kernel_wrapper.cpp/.h，定义 HostCpuKernelWrapperOpV2 及 Initialize，通过 REGISTER_HOST_CPU_OP_BUILDER("HostCpuKernelWrapperOpV2", ...) 注册，将 Compute 转发到 CpuConstantFoldingComputeNew。
• 恢复交付项与构建/测试配套: cmake/package.cmake 与 scripts/package/module/ascend/GeCompilerSo.xml 重新纳入 aicpu_const_folding/libaicpu_const_folding.so，并恢复 constant_folding_ops_stub.cpp 及 tests/engines/cpueng 下的 folding_st.cpp、folding_ut.cpp、aicpu_stub.cpp 测试与桩件。

代码审查

✅ 未发现问题

Thanks for your pull-request.
The full list of commands accepted by me can be found at here.
You can get sig-info at here.
You can self-configure the PR merge rules for this repository. For more details, please refer to Here.
For more, you also can visit HICANN.

PR Approval Progress

✅ Congratulations! All modules have met the lgtm and approve requirements.

Module Approval Details

💡 Tip:

• Committer can comment /approve or /lgtm
• Commenting /approve implies both code review (lgtm) and intent to merge (approve)

CLA Signature Pass

pan-tong, thanks for your pull request. All authors of the commits have signed the CLA. 👍

CLA Signature Pass

pan-tong, thanks for your pull request. All authors of the commits have signed the CLA. 👍

compile

流水线分析报告 — ge run#b596cb88

流水线: https://gitcode.com/CANN/ge/actions/runs/b596cb8826564c1f8c33adad4ebb0df8 | PR#5239 | 2026-09-29 14:29:09 ~ 14:37:11 | 失败 Job 1 个

失败根因分析

[2026-09-29 14:36:00.103+08:00] 2026-09-29 14:36:00,081 - INFO 响应: {"code":"200","msg":"success","data":"https://www.openlibing.com/apps/entryCheckDashCode/MR_a5667f64c2d242a69a670a772bc80ba7/837d1fa17cce0c3f32b515482be0fa78?projectId=300033&codeHostingPlatformFlag=gitcode","uuid":null,"state":"no pass","failData":null,"taskCount":0,"task_id":null}
[2026-09-29 14:36:00.104+08:00] 2026-09-29 14:36:00,083 - INFO [codecheck 2026-09-29 14:36:00] codecheck result: no pass.
[2026-09-29 14:36:00.105+08:00] ::error::Process exited with code 255

由门禁失败定界分析自动生成 | repo: ge | run: b596cb8826564c1f8c33adad4ebb0df8 | PR#5239

compile

compile

/lgtm

/lgtm

/approve