This is a repository for an artifact evaluation of "Tetris: Accelerating Sparse Convolution by Exploiting Memory Reuse on GPU".

ENV

• V100-PCIE-32GB

Build baseline

The evalution depend on Unified-Convolution-Framework and sputnik. To get all source code, run the following code:

git clone --recursive [this-repo-url]

Sputnik

cd /Tetris-artifact-evalution/sputnik
mkdir build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release -DCUDA_ARCHS="70;75"
make -j12
export LD_LIBRARY_PATH=$(pwd)/sputnik/:$LD_LIBRARY_PATH

TACO-UCF

• build TACO-UCF

cd /Tetris-artifact-evalution/Unified-Convolution-Framework
mkdir build && cd build
cmake -DCMAKE_BUILD_TYPE=Release -DCUDA=ON ..
make -j12
export LD_LIBRARY_PATH=$(pwd)/lib/:$LD_LIBRARY_PATH

• build sparse conv

cd /Tetris-artifact-evalution/Unified-Convolution-Framework/benchmark/filter_sparse_convolution/gpu/
make

Install python dependency

cd /Tetris-artifact-evalution/evalution
python3 -m pip install -r requirement.txt 

Test

First, copy the test dataset to TACO-UCF repo

cp -r /Tetris-artifact-evalution/UCF-Dataset  /Tetris-artifact-evalution/Unified-Convolution-Framework/benchmark/filter_sparse_convolution/gpu/

Run the following command to enter specific dir.

cd /Tetris-artifact-evalution/evalution
mkdir ../Tetris/build
mkdir ./figure

To test specific config, run the following command:

./fig6.sh BATCH_SIZE MODEL_NAME

BATCH_SIZE in [1, 4, 8, 16]

MODEL_NAME in [vgg19, mobilnet, yolov3, resnet50]

Example0： vgg19 with batch size 1

./fig6.sh 1 vgg19

Example1: test all models with all batch size

./fig6.sh