ConvWeightCompressFusionPass

Description

For Cube operators, the filter is compressed by inserting a compression operator or sparsified by inserting a 2:4 structured sparsity operator. This pass performs compression or 2:4 structured sparsity based on the user-defined parameters.

Cube operators support Conv2D, FullyConnection, and MatMulV2.

Before:

Restrictions

The first node (Conv2D/FullyConnection/MatMulV2) must meet the following conditions:

• The data type of the input with index 0 must be int8 or uint8.
• AI Core must be supported.
• The value of groups must not be greater than 1.
• Weight compression or 2:4 structured sparsity must be supported.

Availability

The pass effectiveness depends on whether the target platform supports the desired operator type. For details, see .