Closed
[Bug-Report|缺陷反馈]: BatchNorm3DGrad 在 GEIR 图模式下 NCHW/NHWC 用例被内置 V1 推导拦截 #7312
严明奎Add on 27 days agoClosed at 27 days ago
27 days ago Added label: bug-report
27 days ago Close issue
27 days ago Added label: resolved
26 days ago Associated pull request
9 days ago Assign qq_37917228 as responsible person
9 days ago changed custom state from 进行中 to 已完成
Describe the current behavior / 问题描述 (Mandatory / 必填)
BatchNorm3DGrad算子的 tiling 与 binary 配置实际支持NCHW/NHWC/NCDHW/NDHWC四种格式,但在 GE 静态图(GEIR)模式下,NCHW/NHWC用例在 shape 推导阶段即被拦截,走不到本算子的 V2 推导实现。CANN 内置的算子原型库(
libopsproto.so中nn_batch_norm_ops.cc的BatchNorm3DGradInferShape)为同名图算子注册了一份 V1 推导,且仅接受data_format为NDHWC/NCDHW,其余一律报错:GE 的
CallInferFunc在 OpDesc 上存在 V1 推导时优先调用 V1,因此 GEIR 的NCHW/NHWC用例都在内置 V1 推导处被拦截,报 E89999。依据:
norm/batch_norm3_d_grad/op_host/arch35/batch_norm3d_grad_tiling.cpp:247-266(NCHW/NHWC要求 4 维,NCDHW/NDHWC要求 5 维,其余格式才报错)norm/batch_norm3_d_grad/op_host/config/ascend950/batch_norm3_d_grad_binary.json与.../ascend350/...(float16/float32 × NCHW/NHWC/NCDHW/NDHWC)norm/batch_norm3_d_grad/op_host/batch_norm3_d_grad_infershape.cpp(IMPL_OP_INFERSHAPE),未对data_format做字符串校验Environment Information / 环境信息 (Mandatory / 必填)
缺陷所处通路:GE 静态图(GEIR)模式,
AddGraph/RunGraph的 shape 推导阶段 —— 即 GE 的CallInferFunc优先调用 OpDesc 上的 V1 推导注册表的路径。V2(runtime2.0)exe_graph 通路不受影响。具体硬件型号与 CANN 版本未实测,不作填写。
Steps to reproduce the issue / 重现步骤 (Mandatory / 必填)
op::BatchNorm3DGrad节点,data_format设为NCHW(或NHWC)y_backprop/x按 4 维形状(NCHW)设置,scale/reserve_space_1/reserve_space_2按{C}设置graph.SetInputs(...).SetOutputs(...)后AddGraph/RunGraph(参考
norm/batch_norm3_d_grad/examples/test_geir_batch_norm3d_grad.cpp,该示例当前使用NCDHW,改为NCHW即可命中。)Describe the expected behavior / 预期结果 (Mandatory / 必填)
GEIR 图模式下
NCHW/NHWC用例通过 shape 推导,输出形状推导为:与 V2 推导(
op_host/batch_norm3_d_grad_infershape.cpp)语义一致。Related log / screenshot / 日志 / 截图 (Mandatory / 必填)
报错文本(来自内置
libopsproto.so中的 V1 推导实现nn_batch_norm_ops.cc: BatchNorm3DGradInferShape):GE 图编译阶段返回 E89999,推导失败,未进入 tiling。
Special notes for this issue/备注 (Optional / 选填)
本 issue 内容由代码分析得出(内置库报错文本未在本机实测,本机无 CANN 环境)。关联修复 PR 见下方链接。
关联 PR:https://gitcode.com/cann/ops-nn/merge_requests/12696