TbeAippConvReluMaxpoolingFusion
Description
Fuses the Aipp (optional)+Conv2D+Dequant (optional)+ElemWise (optional)+MaxPool (MaxPool/Pooling/MaxPoolv3)+AscendQuant (optional) operators in the following pattern subgraph into one operator.
Restrictions
Fusion can be performed only when the following conditions are met:
- Conv2D: Small channel is enabled. The kernel size must be 3 x 3, 5 x 5, or 7 x 7, strides must be [1, 1] or [2, 2], and cout must be less than or equal to 64.
- MaxPool: strides = [2, 2], ksize = [2, 2]/[3, 3].
- When MaxPool ksize is [2, 2] and the Conv2D input width is greater than 1000, the fusion pattern is disabled.
- When MaxPool ksize is [3, 3] and the Conv2D input width is greater than 800, the fusion pattern is disabled.
- Pooling: strides = [2, 2], window= [2, 2]/[3, 3].
- When Pooling window is [2, 2] and the Conv2D input width is greater than 1000, the fusion pattern is disabled.
- When Pooling window is [3, 3] and the Conv2D input width is greater than 800, the fusion pattern is disabled.
Fusion can be performed only when Maxpoolv3 meets the following conditions:
- soc:
Atlas inference accelerator card - conv2d:
(2) The shape of fmap is [N, 3, 224, 224]. N can be any valid input.
(3) The shape of filter is [N, 3, 7, 7]. N ranges from 1 to 96.
(3) The shape of pads is [3, 3, 3, 3].
(4) The shape of strides is [N, N, 2, 2]. N can be any valid input.
(5) The shape of dilations is [N, N, 1, 1]. N can be any valid input.
(6) The value of groups is 1.
- maxpoolv3:
(2) The shape of strides is [N, N, 2, 2]. N can be any valid input.
(3) The shape of ksize is [N, N, 3, 3]. N can be any valid input.
(4) padding_mode is CALCULATED.
(5) The shape of pads is [1, 1, 1, 1].
(6) global_pooling is false.
(7) ceil_mode is false.
- Aipp: C04 enabled
- Elemwise: Only Relu and LeakyRelu are supported.