ZigZag API

Once ZigZag is available in your Python path, you can import the api in any Python file:

from zigzag import api

get_hardware_performance_zigzag()

This function takes in an workload, a hardware architecture and a mapping file and returns the hardware performance of executing the model’s layers on the architecture under the given mapping constraints.

energy, latency, cme = get_hardware_performance_zigzag(
    workload,
    accelerator,
    mapping,
    opt="latency",
    dump_folder="outputs/{datetime}.json",
    pickle_filename="outputs/list_of_cmes.pickle"
)

The inputs of the function are

• workload: A neural network model defined in ONNX format or ZigZag’s own format.

• accelerator: A high-level hardware architecture description (in ZigZag’s own format).

• mapping: A file that specifies core allocation, spatial mapping (optional), temporal ordering (optional), and memory operand link (in ZigZag’s own format).

• opt (optional): Optimization target. It can be ‘energy’, ‘latency’, or ‘EDP’ (energy-delay-product).

• dump_folder (optional): The name of the result folder.

• pickle_filename (optional): The name of the file which includes all the detailed metadata for analyzing and debugging.

The outputs of the function are

• energy: A number that indicates the overall consumed energy for running the workload on the accelerator in the user-defined optimized way.

• latency: A number that indicates the overall latency (cycle count) for running the workload on the accelerator in the user-defined optimized way.

• cme: A collection of all the detailed cost model evaluation results. “cme” stands for “cost model evaluation”.

Note

We demonstrate how to use this api function in multiple demos.