Concat merging

The RecomposeConcat pass is an ONNX-MLIR optimization pass that simplifies and merges ONNXConcatOp operations to improve model performance and reduce redundant operations.

The pass optimizes Concat operations by:

• Identifies Concat layers with the same axis and combines their inputs into a single Concat node.
• If the Concat has only one input, it directly replaces the Concat with that input.

Input Representation

• Input tensors: {X1,X2,...,Xn}
• Concat axis: A
• Shape of each tensor: (N, C, H, W)

Original Flow (Nested Concat Operations)

• A Concat operation has an input that is itself a Concat:
  Y=Concat(X1,Concat(X2,X3),X4)
• Computational Cost: Each Concat operation involves memory allocations and data movement
  

Optimized Flow (Flattening Nested Concats)

• Instead of a nested structure, the pass flattens the Concat operations:
  Y=Concat(X1,X2,X3,X4)
• Computational Savings: Since Concat now operates in a single step, memory and computation are optimized
  
• The reduction factor in computation can be expressed as: