[BIG tensor]Fix int32 overflow for l1_loss

[Difers]

PR Category

Operator Mechanism

PR Types

Bug fixes

Description

修复paddle.nn.functional.l1_loss底层Reduce Kernel int32溢出问题

Pcard-73145

• 测例

import paddle input = paddle.rand([1, 2147483646], dtype="float16") //接近但小于int32边界 label = paddle.rand([1, 2147483646], dtype="float16") //接近但小于int32边界 l1_loss = paddle.nn.functional.l1_loss(input, label) print(l1_loss) 

• 触发原因
  input.numel() 小于int32 边界 -> 走ReduceAnyKernel indextype int32分支 ->
  for (; input_idx + block_size < bound; input_idx += REDUCE_VEC_SIZE * stride)
  input_idx在循环最后一次可能int32溢出

[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.