[BIG tensor] fix acc error of margin f16

[hxzd5568]

PR Category

Operator Mechanism

PR Types

Improvements

Description

Pcard-67164

1. 改动提升了总体精度，对齐了torch
2. 改动不会影响性能：
   因为改动仅仅影响线性规模的参数，paddle本来 float 16运算过程也做了精度提升，只是有线性个中间结果用float16存储的，所以在保存中间结果的时候又产生了误差。
3. 为什么不在c++层面修改：
   paddle中间过程拆得很碎，拆成了5个中间过程，如果仅对中间结果提升精度，改起来工作量很大。
   (见Paddle/paddle/phi/kernels/gpu/margin_cross_entropy_kernel.cu:19)
4. 实验速度——修复后f16性能略有提升

大tensor 修复后 7.63s Time (%) Total Time (ns) Instances Avg (ns) Med (ns) Min (ns) Max (ns) StdDev (ns) Name -------- --------------- --------- --------------- --------------- ------------- ------------- ------------ ---------------------------------------------------------------------------------------------------- 82.6 6,313,407,732 1 6,313,407,732.0 6,313,407,732.0 6,313,407,732 6,313,407,732 0.0 void phi::HardLabelSoftmaxWithCrossEntropyKernel<float, long>(T1 *, T1 *, const T2 *, int, long, lo… 修复前：8.08s Time (%) Total Time (ns) Instances Avg (ns) Med (ns) Min (ns) Max (ns) StdDev (ns) Name -------- --------------- --------- --------------- --------------- ------------- ------------- ----------- ---------------------------------------------------------------------------------------------------- 84.1 6,873,639,751 1 6,873,639,751.0 6,873,639,751.0 6,873,639,751 6,873,639,751 0.0 void phi::HardLabelSoftmaxWithCrossEntropyKernel<phi::dtype::float16, long>(T1 *, T1 *, const T2 *,… 8.0 649,658,963 

[paddle-bot]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
⚠️ Please upload report for BASE (develop@b4a4977). Learn more about missing BASE report.

Additional details and impacted files

@@ Coverage Diff @@ ## develop #74254 +/- ## =========================================== Coverage ? 100.00% =========================================== Files ? 1 Lines ? 6 Branches ? 0 =========================================== Hits ? 6 Misses ? 0 Partials ? 0 

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[hxzd5568]

/re-run npu

[wanghuancoder]

LGTM