Closed
[Requirement|需求建议]: foreach_reciprocal fp32 倒数精度对齐 IEEE 正确舍入(消除 1 ULP 误差) #8219
wenqi_wuAdd on 17 days agoClosed at 16 days ago
17 days ago Added label: requirement
17 days ago Assign wenqi_wu as responsible person
16 days ago Close issue
16 days ago Added label: resolved


Background(背景信息)
社区问题单 wangyueting/changwei_jiaoyiming#31:torch._foreach_reciprocal 的 fp32 用例(OP-aclnnForeachReciprocal-L0-030/031)精度失败。对比日志显示 NPU 与 golden 的差值全部恰为 1 ULP(输出 3.3e9 量级处 1 ULP=256),third_party(GPU torch)与 golden 逐位相等。经定位:算子 arch35 SIMT 内核以
1.0f/x计算倒数,但 kernel 编译单元未开启高精度编译选项(同族 8 个算子均已开启),默认 fast-math 下除法被编译为近似倒数(rcp 近似 + 牛顿迭代),与 IEEE 正确舍入除法相差 ≤1 ULP。Origin(信息来源)
社区问题单 #31(torchapi 通路,SIMT-LongTails 长尾算子验证);本地复现:ACLNN 直调 70893 元素 fp32 倒数与 CPU IEEE 正确舍入对比,修复前 maxULP=1(约 8%~30% 元素不匹配)。
Benefit / Necessity (价值/作用)
Design(设计方案)
foreach_reciprocal/op_kernel/CMakeLists.txt:arch35 编译单元增加OPTIONS "--cce-use-fast-math=false"(编译期将1.0f/x降级为 IEEE 正确舍入除法)DivAPI,本次未改动(如需同样收紧可另行验证)关联 PR:https://gitcode.com/cann/ops-nn/pull/13740