Skip to content

How to Configure Different Mapping Strategies for Different Types of Convolution Layers (e.g., Standard vs. Depthwise) in ZigZag #116

@ZhangYuanQing99

Description

@ZhangYuanQing99

I’m trying to use ZigZag to evaluate different mapping strategies for different types of convolution layers (e.g., standard convolution vs. depthwise convolution).

From the documentation, I understand that:

1. The mapping definition file allows specifying mappings per operation type (e.g., Conv, ADD, Pooling).

  • name: Add
    spatial_mapping:
    D1:
    - G, 32
    D2:
    - C, 1
    memory_operand_links:
    O: O
    W: I2
    I: I1

  • name: Pooling
    spatial_mapping:
    D1:
    - G, 32
    D2:
    - C, 1
    memory_operand_links:
    O: O
    W: I2
    I: I1

2. It’s also possible to assign custom mappings to individual layers using the name attribute.- name: example_name_of_layer0

  • name: example_name_of_layer0
    spatial_mapping:
    D1:
    - C, 32
    D2:
    - K, 32
    temporal_ordering:
    • [OX, 112] # Innermost loop
    • [OY, 112]
    • [FX, 7]
    • [FY, 7]
    • [K, 2] # Outermost loop
      memory_operand_links:
      O: O
      W: I2
      I: I1

How can I define distinct mappings for different subtypes of convolutions (e.g., standard Conv2D vs. DepthwiseConv2D)?

Thank you for your guidance!

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions