StrideHoistingPass

Description

Inserts the ReadSelect operator based on different graph structures. If Conv2D is matched, the shape and attributes of Conv2D are modified. The ultimate goal is to halve the computation workload.

Scenario 1:

Scenario 2:

Scenario 3:

Scenario 4:

Scenario 5:

Restrictions

• The restrictions on scenarios 1, 2, and 5 are as follows:
  • The length of the path from the node 1 to the node n1 and the length of the path from the node 1 to the node n2 are both less than 10, and there is at least one Conv2D on the path from the node 2 to the node n1 or the path from the node 3 to the node n2. If there is no Conv2D on the path, node 1 must be directly connected to Eltwise/AscendRequantS16. (After the fusion, ReadSelect is inserted between node 1 and Eltwise/AscendRequantS16.) The Conv2D attributes must meet the following requirements:

    Table 1 Conv2D attribute requirements

    H and W Dimension Values of the Second Output Filter	H and W Dimension Values in Stride of Operator Description	H and W Dimension Values in Pads of Operator Description	H and W Dimension Values in Dilations of Operator Description
    3	1	1	1
    5	1	2	1

  • All nodes in the path must be in the trustlist, which includes CONV2D, ELTWISE, RELU, LEAKY_RELU, ASCEND_QUANT, ASCEND_DEQUANT, ASCEND_REQUANT, ASCEND_REQUANTS16, and ASCEND_DEQUANTS16.
• The restrictions on scenarios 3 and 4 are as follows:
  • The first Conv2D in the figure has a single output.
  • The H and W axes of the first input x of the first Conv2D in the figure are static.
• The attributes of the last Conv2D node in all graphs must meet the following requirements:

  Table 2 Last Conv2D attribute requirements

  H and W Dimension Values of the Second Output Filter	H and W Dimension Values in Stride of Operator Description	H and W Dimension Values in Pads of Operator Description	H and W Dimension Values in Dilations of Operator Description
  1	2	0	1

Availability

Atlas 200/300/500 Inference Product

Atlas Training Series Product