[OneDNN][PIR] Fix picodet performance drop in bf16

[zhanglirong1999]

trafficstars

PR Category

Performance Optimization

PR Types

Performance

Description

Scale Op with bias skip bfloat16, since scale with bias add quant/dequant will make result different and lead MultiClassNMSKernel execution time increase 40 times

[paddle-bot[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.