SameInputConv2dPass
Description
Non-quantization scenario: Fuses multiple Conv2D+Relu composites into one Conv2D+Relu+Spilt composite.
Before: 
After: 
Quantization scenario:
Fuses multiple Conv2D+AscendRequant composites into one Conv2D+AscendRequant+Split composite.
Before: 
After: 
Fuses multiple Conv2D+AscendDequant+AscendQuant composites into one Conv2D+AscendDequant+AscendQuant+Split composite.
Before: 
After: 
Restrictions
- Only the static input shape (fmap, filter, bias) is supported.
- Conv2D groups can only be 1.
- The sum of filter batches of the first Conv2D operators (conv2d_0 and conv2d_2 in the figure) must be a multiple of 16. If the data type of fmap is int8, the sum of filter batches must be a multiple of 32.
- filter supports only the const, RequantHostCpuOp, ConvBnFilterHost, and AscendWeightQuant nodes.
- In non-quantization scenarios, the end nodes must be relu or conv2d.
- In quantization scenarios, the end nodes must be dequant, requant, or conv2d.
Availability
Parent topic: Graph Fusion Patterns