SW/HW Prototypes

Benchmark Suites

The most widely used five CNNs (AlexNet, CifarNet, ResNet, SqueezeNet, VGGNet (will be soon uploaded) ) and two RNNs (GRU, LSTM) are written in pure CUDA and OpenCL and tested on GPGPU-Sim, NVIDIA GPUs, and Xilinx FPGAs for computer architecture and NN algorithm research community. Predictions were verified against model files for image recognition (CNNs) and stock trend prediction (RNNs) which were trained by Caffe and Keras. The weight files are included in the benchmark suite. My crazily smart students, Chethan Keshava and Aajna Karki wrote all CUDA code and Spoorthi Mysore Shivakumar and Goutam Madhukeshwar Hegde wrote OpenCL code.

  • The repository is still not complete. But, the correctness of the already uploaded networks are verified so feel free to test them. Please cite GPGPU'19 or ISPASS'19 when you use the benchmark suite.

  • Why Tango? Because Tango music is full of energy and makes you dance! If you don't agree, I recommend you to listen to Libertango and Nightclub 1960 composed by Astor Piazzolla and Tango en Skai written by Roland Dyens. We hope our benchmark suite makes you feel excited. :-)

Simulator Code

This is the simulator code that was used for Locality-aware GPU Register File paper published in CAL in 2019.

  • It was developed based on GPGPU-Sim 3.2.1.

This is the simulator code that was used for GPU Register File Virtualization paper published in MICRO'15.

    • It was developed based on GPGPU-Sim 3.2.1.

    • * Beware, we are not actively maintaining this code. Please cite the above paper when you use this code.

DNN models and datasets

This is the repository of the FCN models and dataset that were used for Thor paper published in IEEE Access in 2023.