[CINN] PostProcess supports x86 composite reduce

[Enigmatisms]

PR Category

CINN

PR Types

Improvements

Description

The following PRs forms the support for x86 end composite reduce (Welford variance & argmin/argmax):

• [CINN] CodeGenLLVM supports composite reduce #72368
• [CINN] PostProcess supports x86 composite reduce #72370
• [CINN] x86 runtime intrinsics for composite reduce #72371

This PR extends the post process module, including the RealizeCompositeReducePass and the Optimize function (minor):

• Instead of creating arithmetic CINN IR nodes (like ir::Max, ir::Add), as in the GPU pass, we replace the node with x86 intrinsic function calls.
• Buffers are correctly re-routed, since x86 passes won't go through scheduling.
• Recompute for LoweredFunc to have updated buffers and their corresponding types.

Pcard-89620

[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.