MatmulToGemmOpFusionPass

Description

Converts the MatMulV3/MatMulV2/MatMul operators that meet the fusion pattern into the GemmV3 operator.

Restrictions

• The Cast node between the MatMul and AssignAdd nodes may not exist. Matching without the Cast node is supported.
• The MatMul type includes MatMul, MatMulV2, and MatMulV3.
• The input dtype of the MatMul node can only be float16, float32, or bfloat16.
• The input dtype of the AssignAdd node can only be float32.
• Disabling this functionality may impact the network accuracy. We recommend not disabling it.

Availability

Atlas 350 Accelerator Card