fix min and max, use loop_kernel when reduction size is larger than 16384

[iclementine]

fix min and max, use loop_kernel when reduction size is larger than 16384

PR Category

Operator

Type of Change

Description

Add an loop-style reduction kernel for min and max. Use the loop-style kernel when reduction size is larger than 16k, and choose a tile size <= 8k to avoid triton tensors larger than triton's limit.

Issue

Progress

• Change is properly reviewed (1 reviewer required, 2 recommended).
• Change is responded to an issue.
• Change is fully covered by a UT.

Performance