Closed
[Documentation|文档反馈]: Xlogy aclnn接口DOUBLE声明与out类型校验按实际支持情况对齐 #4550
zhaoxinmengAdd on  14 days agoClosed at  14 days ago
zhaoxinmengCollaborators
14 days ago Create

一、文档链接 (必填)

  • math/xlogy/docs/aclnnXLogYScalarOther&aclnnInplaceXLogYScalarOther.md(含_en)
  • math/xlogy/docs/aclnnXLogYScalarSelf.md(含_en)
  • math/xlogy/docs/aclnnXLogYTensor&aclnnInplaceXLogYTensor.md(含_en)
  • math/xlogy/op_api/aclnn_x_log_y_tensor.cpp

二、问题片段 (必填)

  1. aclnnXLogYScalarOther(含Inplace)的标量参数other文档声明9种,漏列DOUBLE(实现按SCALAR支持列表10种校验,上板实测接受且计算正确),与同族ScalarSelf的self标量(10种含DOUBLE)口径不一致。
  2. 标量接口(ScalarSelf/ScalarOther)的out文档声明3种浮点,实测out=DOUBLE被接受且值正确(fp32域计算结果)。
  3. aclnnXLogYTensor第一段接口未校验out数据类型:列表外dtype(如UINT16/UINT32/UINT64/DOUBLE)依赖运行时l0op::Cast静默承接(实测out=UINT16可执行成功且值正确),与接口文档out声明口径不符。

💡 修复内容

按SE确认口径(以真实代码支持情况为准):

  • ScalarOther(含Inplace)的标量参数other补回DOUBLE(10种),与ScalarSelf对齐
  • 标量接口的out补DOUBLE(3种浮点→4种)
  • Tensor接口的out声明按实测扩充为15种(补DOUBLE、UINT16、UINT32、UINT64、COMPLEX64、COMPLEX128):第一段接口对out不做列表校验,运行时l0op::Cast承接全部15种且值比对通过(经权衡不改代码,资料按真实行为声明)
  • 精度口径:DOUBLE标量/输出按FLOAT精度计算(实现收窄FLOAT后经Cast扩展,kernel无FLOAT64 binary,与fp64域计算结果差≤1 ulp(fp32))

实测验证(Ascend 950PR,TTK):

  • Tensor接口out类型支持矩阵(15种):15/15全PASS(第一段Success+执行成功+值比对通过,含COMPLEX64/COMPLEX128);L0功能精度回归8/8全PASS
  • 标量接口other=DOUBLE/out=DOUBLE:实测接受且计算正确

关联PR:https://gitcode.com/cann/ops-math/pull/7143

likedislike
ZzhaoxinmengCollaborators
14 days ago Added label: documentation
zhaoxinmengCollaborators
14 days ago Comments:

/assign @zhaoxinmeng

likedislike
CANN-robotCANN-robotCollaborators
14 days ago Assign zhaoxinmeng as responsible person
ZzhaoxinmengCollaborators
14 days ago Modify description of issue
CANN-robotCANN-robotCollaborators
14 days ago Close issue
CANN-robotCANN-robotCollaborators
13 days ago Added label: resolved