ZConfusionSoftmaxGradFusionPass
Description
Fuses the Mul, ReduceSumD/ReduceSum (supported only by Atlas 350 Accelerator Card), and Sub operators that fit the graph fusion pattern into the ConfusionSoftmaxGrad operator.

Restrictions
- The fusion pattern does not take effect in the dynamic shape scenario.
- The fusion pattern does not take effect when the two inputs of the Mul operator have different shapes.
- The fusion pattern does not take effect when the input0 of the Mul operator and the input0 of the Sub operator are from different nodes.
- The fusion pattern does not take effect when input0 is not the first input of the Sub operator.
- The fusion pattern does not take effect when the axis attribute of ReduceSumD/ReduceSum does not correspond to the last axis index of the input shape.
- The fusion pattern does not take effect when the last axis of the input shape for ReduceSumD/ReduceSum is greater than 30000.
- The fusion pattern does not take effect when the data type is not supported by the fused operator.
- The data type and format of input0 must be the same as those of input1.
- The fusion pattern does not take effect if the reduce axis is set to 1 and the AReduceSumFusionPass fusion pattern is enabled.
- For Atlas 350 Accelerator Card, if keep_dims of ReduceSum is set to False, the fusion pattern does not take effect.
- For Atlas 350 Accelerator Card, if the input axis of ReduceSum is not a Const node, the fusion pattern does not take effect.
Availability
Atlas 350 Accelerator Card
Parent topic: Graph Fusion Patterns