MatmulReduceSumUbFusion
Description
Performs UB fusion on BatchMatMul and ReduceSum in the following pattern subgraph.

Restrictions
- The input of BatchMatMul is not 1, and the output is 1D.
- The output of ReduceSum is of type float32, and keep_dim is false.
- BatchMatMul and BatchMatMulV2 are supported.
- The shape of the MatMul input data cannot exceed three dimensions. In addition, dimension 0 of the input data cannot be 1 and cannot exceed the maximum value of uint16_t.
Availability
Parent topic: SpaceToBatchConv2dBatchToSpacePass