ZigZag: A Joint Architecture-Mapping Design Space Exploration Framework for DNN Accelerators

Paper

Old version: https://arxiv.org/abs/2007.11360

New version: https://ieeexplore.ieee.org/document/9360462

Functions of ZigZag

Supported Mode	Function	Memory Scheme Search	Spatial Unrolling Search	Temporal Mapping Search	Cost Estimation
1	Hardware Cost Evaluation	No	No	No	Yes
2	Find the best Temporal Mapping (minimize energy or maximize throughput) for a single NN layer	No	No	Yes	Yes
3	Find the best Spatial Unrolling and the best Temporal Mapping for a single NN layer	No	Yes	Yes	Yes
4	Find the best Memory Scheme with fixed Spatial Unrolling for a single NN layer	Yes	No	Yes	Yes
5	Find the best Memory Scheme with the best Spatial Unrolling for a single NN layer	Yes	Yes	Yes	Yes
6	Find the best Spatial Unrolling for multiple NN layers (could be a complete NN or multiple NNs)	No	Yes	Yes	Yes
7	Find the best Memory Scheme with fixed Spatial Unrolling for multiple NN layers (could be a complete NN or multiple NNs)	Yes	No	Yes	Yes
8	Find the best Memory Scheme with the best Spatial Unrolling for multiple NN layers (could be a complete NN or multiple NNs)	Yes	Yes	Yes	Yes

Quickstart

To run the framework

python3 top_module.py \
--set <path_to_settings_file> \
--map <path_to_mapping_file> \
--mempool <path_to_mempool_file> \
--arch <path_to_arch_file> \

Examples

A few input setting files are contained in the inputs folder.

Single cost estimation

In the example provided a single cost estimation is carried out for the inference of CONV4 of AlexNet on Eyeriss.

In the settings file (inputs/settings.yaml) the architecture and the mapping of the dataflow are fixed. (fixed_architecture, fixed_spatial_mapping and fixed_temporal_mapping are all set to True).

The architecture specs are defined in inputs/architecture.yaml while the mapping specs are define in inputs/mapping.yaml. For more info on how these specifications are set, refer to Input setting parameters.

The cost estimation can be run with: python3 top_module.py --arch ./inputs/architecture.yaml --map ./inputs/mapping.yaml --set ./inputs/settings.yaml --mempool ./inputs/memory_pool.yaml

Temporal mapping exploration

A temporal mapping exploration can be carried out on the same architecture with the same workload by setting fixed_temporal_mapping to False in the settings file.

If the temporal mapping exploration is enabled, a search method must be specified. The search method (exhaustive, heuristic_v1, heuristic_v2, or iterative) can be set in the settings file.

The temporal mapping exploration can be then run with: python3 top_module.py --arch ./inputs/architecture.yaml --map ./inputs/mapping.yaml --set ./inputs/settings.yaml --mempool ./inputs/memory_pool.yaml

Spatial unrolling exploration

Beside the temporal mapping exploration, spatial unrolling exploration can be carried out by setting the fixed_spatial_mapping to False as well.

If the spatial unrolling exploration is enabled, a search method must be specified. The search method (exhaustive, heuristic_v1, heuristic_v2, or hint_driven) can be set in the settings file.

A MAC array utilization threshold of >0.75 is suggested for reducing the exploration space. It can be specified in the settings file (spatial_utilization_threshold)

The spatial unrolling exploration can be then run with: python3 top_module.py --arch ./inputs/architecture.yaml --map ./inputs/mapping.yaml --set ./inputs/settings.yaml --mempool ./inputs/memory_pool.yaml

Architecture exploration

A basic architecture exploration run can be started by setting the fixed_architecture parameter to False in the settings file.

Ther architecture exploration can be then run with: python3 top_module.py --arch ./inputs/architecture.yaml --map ./inputs/mapping.yaml --set ./inputs/settings.yaml --mempool ./inputs/memory_pool_exploration.yaml

Input settings parameters

Please refer to Input setting parameters

Output file data format

Please refer to Example result files

Console information

While the tool is running, it prints some useful information on the console. Understand this information helps user to to better understand and control the DSE flow.

Please refer to Console information

---

The research team is continuing building and polishing ZigZag. We welcome any comment, discussion, and contribution from the community.