Thanks for sending an requirement! Please fill in the following template to help quickly solve your problem.
cube算子团队
提升Matmul算子FP2场景的精度
L0C累加位宽与atomic累加位宽存在差距
综合考虑性能精度,splitSingleCoreKBase大小为: outer_cnt / innerr_cnt <= 24 innerr_cnt = splitSingleCoreKBase/8 out_cnt = K/splitSingleCoreKBase
推导出: (K/splitSingleCoreKBase) / (splitSingleCoreKBase/8) <= 24 => splitSingleCoreKBase >= sqrt(K/3) splitSingleCoreKBase 满足最小16对齐
Thanks for sending an requirement! Please fill in the following template to help quickly solve your problem.
Backgroud(背景信息)
Origin(信息来源)
cube算子团队
Benefit / Necessity (价值/作用)
提升Matmul算子FP2场景的精度
Design(设计方案)
L0C累加位宽与atomic累加位宽存在差距
综合考虑性能精度,splitSingleCoreKBase大小为:
outer_cnt / innerr_cnt <= 24
innerr_cnt = splitSingleCoreKBase/8
out_cnt = K/splitSingleCoreKBase
推导出:
(K/splitSingleCoreKBase) / (splitSingleCoreKBase/8) <= 24 => splitSingleCoreKBase >= sqrt(K/3)
splitSingleCoreKBase 满足最小16对齐