SplitConvConcatFusionPass

Description

The following scenarios are fused:

Scenario 1

Scenario 2

Scenario 3

Restrictions

• The fusion matches ASplitConv2dConcatPass first. This FusionPass is matched only when the ASplitConv2dConcatPass restrictions are not met.
• ConcatD/ConcatV2D has at least two different input nodes..
• The dim axis of ConcatD/ConcatV2D and SplitD/SplitVD must be the C axis, and the C axis must be aligned with dtype.
  • When the original dtype is fp16 or float, the value of dim C must be a multiple of 16.
  • If the original dtype is int8, the value of dim C must be a multiple of 32.
  • If the original dtype is int4, the value of dim C must be a multiple of 64..
• All ConcatD/ConcatV2D inputs are 4-dimensional and the input format is NCHW or NHWC.
• SplitD/SplitVD has at least two different output nodes..
• The number of ConcatD/ConcatV2D inputs is the same as the number of SplitD/SplitVD outputs..
• The AscendQuant operators connected to SplitD/SplitVD must have the same attributes including scale, offset, and sqrt_mode.

Availability

Atlas 200/300/500 Inference Product

Atlas Training Series Product