[PHI] Fix paddle.kthvalue for big tensor

[cangtianhuang]

PR Category

Operator Mechanism

PR Types

Bug fixes

Description

Bug Fixes:

1. Add typename IndexType to distinguish index types
2. Change shared_mem to dynamic allocation, instead of hardcoding
3. Set grid_dim, block_dim to avoid exceeding
4. Resolve write race condition for foundKValue by adopting CudaAtomicMin
5. Fix grad_kernel int64_t handling

Tests:

The accuracy diff is due for multiple identical kth values, resulting in different index positions.

After comparing only the first value, all accuracy tests pass

TODO:

• 1. Support int64_t k param in API signature
• 2. Replace shared_mem in RadixTopK with dynamic allocation

Pcard-85711

[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.

[wanghuancoder]

LGTM