Skip to content

Repository files navigation

CUDA Kernels for Convolutional Neural Networks (CNNs)

This project is based on an assignment for UCLA's graduate-level course -- Current Topics in Computer Science: System Design/Architecture: Learning Machines.

The goal is to design and implement convolution and classification kernels in CUDA that improve on CPU and GPU baselines (see DianNao, and Baidu's DeepBench, which leverages cuDNN acceleration).

Our kernels improve on DianNao by three orders of magnitude, but fall short of cuDNN by roughly the same.

About

Custom CUDA kernels for running convolutional neural network operations -- i.e. convolution and general matrix multiply

Resources

Stars

1 star

Watchers

2 watching

Forks

Releases

Packages

Contributors

Languages