TbeFullyconnectionElemwiseDequantFusionPass
Description
Performs UB fusion on the FullyConnection/MatMul/MatMulV2/BatchMatMul/BatchMatMulV2+ElemWise+AscendQuant+AscendDequant nodes in the subgraphs that meet the following patterns.
The dotted boxes in the fifth pattern indicate that these nodes may not be matched.
or
or
or
or
Constraints
- The ElemWise2 node supports only Elu, LeakyRelu, Gelu, Softsign, Relu6, Relu, Softplus, Sigmoid, Tanh, Selu, GeluGrad, Add, AddN, FastGelu, FastGeluV2, FastGeluGrad, Eltwise, PRelu, Mul, Muls, Power, Relu6D, and TanhGrad.
- Dynamic shapes are not supported.
- For Matmul, dequant, GELU, and quant operators, the ElemWise1 node must be a Gelu node, and the data type must be fp32.
- If ElemWise2 is not empty, ElemWise1 must be Relu, LeakyRelu, Add, Muls, or AddN, and ElemWise2 must be Relu6.
- When ElemWise1 is Add, the input node of Add must be 2, the output node must be 1, and the previous node must be FullyConnection.
- When ElemWise1 is LeakyRelu, negative_slope is required. The absolute value of negative_slope must be greater than 1.19209e-07.
- MatMul+ElemWise1 cannot be AddN or Mul.
- For BatchMatMul/BatchMatMulV2, ElemWise2 cannot be Add or Relu.
- When elemwise_node type is add, the output shape of fc cannot be smaller than the input shape of add.
Applicable Products
Parent topic: UB Fusion Patterns