OCPA

This repository is the implementation of Accelerating Convolutional Neural Network by Exploiting Sparsity on GPUs

Requirements

+ Ubuntu 18.04
+ cuda == 10.2
+ cuDNN == 7.6

• CUDA == 12.2
• cuDNN == 8.8+

For using CMake and C++ you can use clang, llvm library can be used in addition:

• llvm-16
• clang-16+

Experiments

• Vgg-16

running ECR for convolution layer

cd OCPA/ECR/ECR/time_vgg
nvcc batchedECR.cu -o batchedECR.out
./batchedECR.out 32

running PECR for convolution+pooling layer

cd OCPA/PECR/pecr/time_vgg
nvcc batchedPECR.cu -o batchedPECR.out
./batchedPECR.out 32

running cudnn(using tensor core) for convolution layer

cd OCPA/ECR/cudnn/time_vgg
make
./cudnn 32

running cudnn(using tensor core) for convolution+pooling layer

cd OCPA/PECR/cudnn/time_vgg
make
./cudnn 32

• Resnet-50

running ECR for convolution layer

cd OCPA/ECR/ECR/time_resnet
nvcc batchedECR.cu -o batchedECR.out
./batchedECR.out 32

running PECR for convolution+pooling layer

cd OCPA/PECR/pecr/time_resnet
nvcc batchedPECR.cu -o batchedPECR.out
./batchedPECR.out 32

running cudnn(using tensor core) for convolution layer

cd OCPA/ECR/cudnn/time_resnet
make
./cudnn 32

running cudnn(using tensor core) for convolution+pooling layer

cd OCPA/PECR/pecr/time_resnet
make
./cudnn 32

We can get the running time of other algorithms by analogy with the above methods using cuDNN.

Expected Result

The Vgg-16 and Resnet-50 speedup effects can be obtained by running programs under the folder OCPA/speedup. As shown in the following figure：