MatMulAlignInputsFusionPass

Description

When the inner axis of the input shape of the matmul operator is not 512-byte aligned, the MTE efficiency is low and the performance is poor. This graph fusion aligns the input shape of matmul to solve the performance problem.

Restrictions

Applies only to the static scenario where the input DType is Float32 and the input does not have bias.

This non-common fusion is performed only in some cases.

Availability

Atlas A2 training product/Atlas A2 inference product

Atlas A3 training product/Atlas A3 inference product