ConvFormatRefreshFusionPass

Description

For the structure below, configures the shape and format of the CONV2D operator's Input1 to match those of Output.

Constraints

The fusion pattern takes effect when the following conditions are met:

• The original format of Input1 must be consistent with that of Output, or the original shape dimensions of Input1 must be consistent with those of Output.
• If the output format is 5HD and a C0 value exists, the C0 value of Input1 for the convolution operator must be consistent with that of Output.
• When the format of Input1 is not consistent with that of Output, the output format is restricted to NCHW, NHWC, HWCN, CHWN, or NC1HWC0.

Applicable Products

Atlas 200I/500 A2 inference products

Atlas inference products

Atlas training products

Atlas A2 training products/Atlas A2 inference products

Atlas A3 training products/Atlas A3 inference products

Ascend 950PR/Ascend 950DT