LayoutLoop

LayoutLoop

1.

LayoutLoop is a tool based on TimeLoop. It integrates the functionalities of more accurate layout-based memory modeling.

The key contributions of SquareLoop over previous tools are:

• realistic layout-based memory model utilizing accurate dataspace-wise evaluation
• introduction of physical ranks, allowing for independent per-dataspace layout and AuthBlock specification
• Layout-Mapping co-search algorithm

2. Setup

To avoid the tedious dependency, we offer the docker with all dependencies and code being setup.

2.0 Files Overview

We use the following files in the experiments:

• Architecture
  • SIGMA (vector256, full-flege flexible accelerator)
    • benchmarks/arch_designs/vector_256.yaml
  • SIMBA (reconfigurable systolic array)
    • benchmarks/arch_designs/simba_like.yaml
    • benchmarks/arch_designs/components/*
    • benchmarks/arch_designs/constraints/*
  • Edge-TPU (systolic)
    • benchmarks/arch_designs/vector_256.yaml
    • benchmarks/arch_designs/systolic_constraint/mapspace_XY_OS.yaml
    • benchmarks/arch_designs/systolic_constraint_depthwise/mapspace_XY_OS.yaml
  • Eyeriss (eyeriss)
    • benchmarks/arch_designs/eyeriss_like/arch/eyeriss_like.yaml
    • benchmarks/arch_designs/eyeriss_like/arch/components/*
    • benchmarks/arch_designs/eyeriss_like/constraints/* (constraint for convolution workload only)
    • benchmarks/arch_designs/eyeriss_like/constraints_depthwise/* (constraint for depth-wise convolution workload only)
• Workloads
  • ResNet18
    • benchmarks/layer_shapes/resnet18/*
  • ResNet50
    • benchmarks/layer_shapes/resnet50/*
  • MobileNetV3 (mobv3)
    • benchmarks/layer_shapes/mobv3/*
  • bert
    • benchmarks/layer_shapes/bert/*
  • bert_conv (converted matrix multiplication as the form of convolution)
    • benchmarks/layer_shapes/bert_conv/*
  • vgg small
    • benchmarks/layer_shapes/vgg01/*
  • vgg large
    • benchmarks/layer_shapes/vgg02/*
  • AlexNet
    • benchmarks/layer_shapes/AlexNet/*
• Mapper
  • benchmarks/mapper/mapper.yaml

2.1 Software Dependency -- Docker installation

sudo apt-get install docker-ce docker-ce-cli containerd.io docker-buildx-plugin docker-compose-plugin

2.2 Download and Setup prebuilt docker.

Steps: download the docker link and install it

docker image ls
docker load -i feather_layoutloop_docker.tar.gz 

View the image name from the all available docker images.

3. Experiment: Launch the run for different accelerators setup (Optional, > 24 hours)

#docker run -it <docker_img_name>
docker run -it feather_layoutloop

When inside the docker

pip install torch torchlens pyyaml torchvision pandas
git clone <provided_url>
#e.g. git clone https://github.com/maeri-project/FEATHER.git

cd FEATHER/LayoutLoop/layoutloop
scons -j<number_of_available_threads>
cd FEATHER/LayoutLoop/configurations
make clean
make dse  # launch dataflow design space exploration for various architectures under ResNet-18, MobileNet-V3 and Bert -- using layoutloop based precise memory modeling

The old pre-searched results are listed in the pre_run_results, and the collected results are listed in the function named figure13() in FEATHER/results_generation.py.

4. Results Analysis

4.1 Pre-run results analysis (Mandatory, just reading the prerun-results, take ~5 minutes)

1. All pre-run results are sitting in the folder FEATHER/LayoutLoop/pre_run_results

└── results_precise_layout_modeling

2. In each folder, there are 4 different csv files

├──utilization.csv: the average computation utilization of searched dataflow under designated layout (e.g. 1 mean 100% utilization)
├──cycle.csv: the overall clock cycle of processing given workload levearging the searched dataflow under designated layout (e.g. 452313.00 mean 452313 clock cycles)
└──pj_commpute.csv: computation energy efficiency of processing given workload levearging the searched dataflow under designated layout (e.g 2.17 mean 2.17 pJ/MAC)

The number of row in each csv file is the total number of layer for given workloads. For ease of reading searched results, we also provide the interleave_layoutloop_search.csv to merge all above four files together.

• column 1: utilization.csv
• column 2: cycle.csv
• column 3: pj_commpute.csv

3. The searched dataflow is located at the mapping_search directory. The number of row in each csv file is the total number of layer for given workloads, and the index indicating the layer index.

4.2 New results analysis (Optional, only needed if you run 2,3 above, take ~5 minutes )

Finishing experiments, the results are stored in the configurations/results, while the name pattern of the results is shown as followes

{design_name}_interleave_layoutloop_search.csv

where,

• workloads are "resnet50" (53 layers), "mobv3" (62 layers), "bert" (3 layers). In total, you will see 118 layers (rows in the csv).
• design_name could be "gemmini", "eyeriss", "sigma", "simba", "medusa", "systolic_array".
• layout_policy could be "SRCQPMNHW_Cx32", "SRCQPMNHW_Hx32", "SRCQPMNHW_Mx32", "SRCQPMNHW_Wx32", "SRCQPMNHW_Cx4Hx8", "SRCQPMNHW_Cx8Hx4", "SRCQPMNHW_Cx8Wx4", "SRCQPMNHW_Wx8Hx4", "SRCQPMNHW_Cx8Wx2Hx2".
• column 1: utilization
• column 2: cycle
• column 3: pj_commpute

Have fun! Enjoy XD