Merged
fix(optim): InplaceApplyCenteredRMSProp/InplaceApplyPowerSign 合规质量整改 #14077
tan_xinAdd on 12 days ago
fix(optim): InplaceApplyCenteredRMSProp/InplaceApplyPowerSign 合规质量整改 #14077
Merged
tan_xinAdd on 12 days ago
tan_xin
tan_xinCollaborators
12 days ago

描述

本 PR 对 InplaceApplyCenteredRMSProp 与 InplaceApplyPowerSign 两个算子进行合规质量整改,修复质量规则/静态检查发现的问题。

InplaceApplyCenteredRMSProp

  1. 标量输入校验(op_host/arch35/inplace_apply_centered_rms_prop_tiling_arch35.cpp)
    • 新增 lr/rho/momentum/epsilon(输入索引 4-7)校验:rank 必须 ≤1、元素数必须为 1、dtype 必须与 var 一致,否则返回 GRAPH_FAILED。
    • 修复此前无校验导致的"多元素标量静默取首元素、混合 dtype 在 kernel GM 回退路径按 var dtype 误读"的静默错误。
  2. TQUE 缓冲合规(op_kernel/arch35/inplace_apply_centered_rms_prop.h)
    • VECIN 5 个输入队列由双缓冲改为单缓冲(新增 INPUT_BUFFER_NUM=1),同位置峰值 5 块 ≤ A5 上限 8;VECOUT 4 个输出队列保持双缓冲(4×2=8 恰达上限)。
  3. golden 精度整改(tests/assets/golden.py)
    • 显式登记 __spec__;参考数学改由 Torch 基础算子组合 + float64 高精度域独立计算并 round 到输入 dtype,避免 |grad|>sqrt(FLT_MAX) 时 grad² 上溢与 kernel 同错而被比较掩盖。
    • 新增 third_party 三方交叉校验:torch 独立组合 + tf.raw_ops.ResourceApplyCenteredRMSProp;tolerance 采用 cross_check L2(最严档),触发 need_3party GPU 三方比对。
  4. README 顺序与格式:use_locking 属性行移至输出参数之后;调用说明表头统一为"调用方式/样例代码"、"图模式"→"GE图模式"。CMakeLists 许可头规范化。

InplaceApplyPowerSign

  1. 标量 shape 约束与说明:lr/logbase/sign_decay/beta 由"shape=[1]"更新为"shape=[1] 或 0-D([])";补充动态 Shape/Rank(GE图模式)支持说明(输出 shape 与对应输入一致,rank ≤8)。
  2. 示例资源与错误处理(examples/arch35/test_geir_inplace_apply_power_sign.cpp):GenOnesDataFloat32/GenOnesData 增加分配失败判空与 delete[] 释放;Session::AddGraph 失败分支补充错误处理与 delete session;结束处 delete session。
  3. README/许可头:use_locking 行位置调整、调用说明表头统一、CMakeLists 许可文案修正。

关联的Issue

测试

  • 新增/更新 UT 与 golden 校验覆盖上述改动:inplace_apply_centered_rms_prop golden 三方交叉校验(cross_check L2)及 kernel/GEIR 通路;inplace_apply_power_sign 示例资源释放与错误处理路径。

文档更新

  • optim/inplace_apply_centered_rms_prop/README.md、README_en.md
  • optim/inplace_apply_power_sign/README.md、README_en.md

类型标签

AI/Agent生成声明

likedislike
Pull Request has been successfully merged by CANN-robot.
(Thank you for tan_xin's contribution.)
tan_xintan_xinCollaborators
12 days ago Created pull request, commit fe66aecc
atomgit-bot
atomgit-bot
12 days ago Comments:

变更摘要

本 PR 针对 heaviside 算子完善健壮性与接口描述一致性:在 op_host 的 tiling 与 infershape 阶段补齐空指针、dtype 一致性和溢出边界校验,修正 CeilA2B 的潜在溢出;同步修正 aclnn_heaviside.h 的参数注释,补全 examples 示例程序在各失败分支的资源释放,并把 UT 中的 elementNum 字段宽度对齐为 int64_t;其余多为行尾换行补齐与测试脚本格式清理。

主要改动

  • RunBigKernelTiling 校验增强:新增对 GetInputDesc、GetInputShape、GetPlatformInfo、GetWorkspaceSizes、GetRawTilingData 的空指针判断,并在 values 与 input 的 dtype 不一致、GetCoreNumAiv 返回 0、GetShapeSize 返回 int64_t 最小值(溢出哨兵)时返回 ge::GRAPH_FAILED,避免后续解引用或错误分核。
  • CeilA2B 溢出修复:将 (a + b - 1) / b 改为 (a - 1) / b + 1,并对 b == 0、a <= 0 单独返回,避免 a 接近 INT64_MAX 时有符号加法溢出。
  • InferDataType4Heaviside 数据类型校验:新增读取 values 的 dtype,当 input/values 为 DT_UNDEFINED 或两者不一致时报错并返回 GRAPH_FAILED,否则以 input 的 dtype 设置输出类型。
  • aclnn_heaviside 接口注释修正:out 参数标注由 @param [in] 改为 @param [out],workspaceSize 说明中的第二段接口来源由 aclnnForeachAbsGetWorkspaceSize 更正为 aclnnHeavisideGetWorkspaceSize。
  • 示例与 UT 配套调整:test_aclnn_heaviside.cpp 在 aclrtSetDevice、aclrtCreateStream、aclrtMemcpy、CreateAclTensor 等失败分支补充 aclrtFree/aclDestroyTensor/aclrtResetDevice/aclFinalize 释放;heaviside_tiling_def.h 中 HeavisideTilingData::elementNum 由 int32_t 改为 int64_t 以与 tiling 数据保持一致。

🧩 本次 Review 已使用自定义 Skill:

likedislike
Incorrect?
atomgit-bot
atomgit-bot
12 days ago Comments:

代码审查

审查结论

本 PR 主体是「空指针/返回值加固 + 行尾与格式清理」,核心逻辑改动集中在 heaviside_tiling.cpp(CeilA2B 防溢出重写、dtype/shape/platform/workspace 校验)、heaviside_infershape.cpp(dtype 一致性校验)与 UT 侧 heaviside_tiling_def.h(elementNum 由 int32_t 改为 int64_t,与 host TILING_DATA_FIELD_DEF(int64_t, elementNum) 对齐,属于正确修复)。未发现 P0/P1 级缺陷:CeilA2B 新写法的数学等价性成立(b 恒为 8192 常量,a>0 时 (a-1)/b+1 == ceil(a/b)),kernel 侧 elementNum 本身就是 int64_t,UT 结构体改动方向正确且 sizeof 布局与 host 一致。

发现 5 个问题,均为中低优先级:1 个 P2(kernel UT 精度校验恒通过,测试失效风险),4 个 P3(可选改进/无效校验/测试覆盖缺口)。

各文件审查情况

  • activation/heaviside/examples/test_aclnn_heaviside.cpp:no issues(错误路径补 aclFinalize/aclrtFree 正确,无双重释放;未补 aclrtDestroyStream 属示例级可忽略,aclrtResetDevice+aclFinalize 已释放资源)
  • activation/heaviside/op_api/aclnn_heaviside.h:no issues(out 改 [out]、aclnnForeachAbs→aclnnHeaviside 文案修正正确)
  • activation/heaviside/op_graph/CMakeLists.txt:no issues(仅空白/行尾)
  • activation/heaviside/op_graph/heaviside_proto.h:no issues(仅行尾)
  • activation/heaviside/op_host/config/ascend910_93/heaviside_binary.json:no issues(仅行尾,key 无拼写问题)
  • activation/heaviside/op_host/config/ascend910b/heaviside_binary.json:no issues(同上)
  • activation/heaviside/op_host/config/ascend950/heaviside_binary.json:no issues(同上)
  • activation/heaviside/op_host/config/kirin9030/heaviside_binary.json:no issues(同上)
  • activation/heaviside/op_host/config/kirinx90/heaviside_binary.json:no issues(同上)
  • activation/heaviside/op_host/heaviside_infershape.cpp:1 个 P3(新增 dtype 一致性分支无有效 UT 覆盖)
  • activation/heaviside/op_host/heaviside_tiling.cpp:2 个 P3(溢出哨兵值错误使校验失效;SaveToBuffer 返回值未校验)
  • activation/heaviside/op_host/heaviside_tiling.h:no issues(仅行尾)
  • activation/heaviside/op_kernel/heaviside.cpp:no issues(仅行尾)
  • activation/heaviside/op_kernel/heaviside.h:no issues(仅行尾)
  • activation/heaviside/tests/CMakeLists.txt:no issues(仅空白/行尾;file(GLOB) 为存量模式)
  • activation/heaviside/tests/st/aclnnHeaviside/atk_aclnnHeaviside.json:no issues(单行文件仅行尾变更;dtype/键名 input/values/output 一致)
  • activation/heaviside/tests/ut/CMakeLists.txt:no issues(仅空白/行尾)
  • activation/heaviside/tests/ut/op_host/CMakeLists.txt:no issues(仅空白/行尾)
  • activation/heaviside/tests/ut/op_host/op_api/CMakeLists.txt:no issues(仅空白)
  • activation/heaviside/tests/ut/op_host/test_heaviside_infershape.cpp:no issues(仅行尾;输入名问题随 infershape 变更合并上报)
  • activation/heaviside/tests/ut/op_kernel/CMakeLists.txt:no issues(仅空白)
  • activation/heaviside/tests/ut/op_kernel/heaviside_data/compare_data.py:2 个问题(P2 恒通过、P3 缺 bf16 解析)
  • activation/heaviside/tests/ut/op_kernel/heaviside_data/gen_data.py:no issues(删无效 import、补尾逗号均正确,os.system 与 output_size 未用为存量问题)
  • activation/heaviside/tests/ut/op_kernel/heaviside_tiling_def.h:no issues(elementNum 改 int64_t 与 host tiling 定义一致,属修正)
  • activation/heaviside/tests/ut/op_kernel/test_heaviside.cpp:no issues(仅行尾)

整体风险判断:改动整体是防御性加固,方向正确、回归风险低;主要隐患在测试侧——kernel 精度 UT 实际上永远不会失败(P2),使本次 tiling/类型相关改动的回归验证形同虚设,建议优先修复。

类型 数量
🔴 阻塞 0
🟡 建议 1

💬 仅评论


🧩 本次 Review 已使用自定义 Skill:

likedislike
Incorrect?
CANN-robotCANN-robotCollaborators
12 days ago Added label: cann-cla/yes
CANN-robot
CANN-robotCollaborators
12 days ago Comments:

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

module lgtm status approve status
*/*/README.md ✅ 汤平川, 上官秦南, 陈娇 (3/2) ✅ 陈娇 (1/1)
optim ✅ 汤平川, 上官秦南 (2/2) ✅ 汤平川 (1/1)

💡 Tip:

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

CLA Signature Pass

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

likedislike
52 messages are collapsed here。 See more
TangPC
TangPCCollaborators
10 days ago Comments:

/lgtm
/approve

likedislike
CANN-robotCANN-robotCollaborators
10 days ago Added label: lgtmapproved
CANN-robotCANN-robotCollaborators
10 days ago Closed associated issue
CANN-robotCANN-robotCollaborators
10 days ago Merged pull request
CANN-robot
CANN-robotCollaborators
10 days ago Comments:

Pull Request 已合并或已关闭。

If you want to solve this problem, you can click here to do it in the FAQs.

likedislike