TensorScatterUpdateFusionPass

Description

Splits the TensorScatterUpdate operator that fits the graph fusion pattern into the TensorMove and ScatterNdUpdate operators when the input and output data types are not bool or int32.

Restrictions

• This fusion pattern does not take effect when the data type of input x is bool or int32.
• This fusion pattern cannot be disabled.

For Atlas 350 Accelerator Cards, there are additional restrictions.

• The ScatterNdUpdate operator AI Core implementation supports the following data types: int64, int8, float32, float16, bfloat16, and bool. Therefore, after the fusion is successful, only these six data types execute the AI Core, and other types execute the AI CPU.
• This fusion pattern does not take effect when the data type is String or complex128, since these types are not supported by TensorMove or ScatterNdUpdate.

Availability

Atlas A2 training product/Atlas A2 inference product