Open CUDA convolution neural network inference implementation
Abstract
Abstract This work presents an open, efficient (fast) CUDA convolution neural network inference implementation specialized in some layers of popular nets like ResNet, VGG, and GoogLeNet. The proposed algorithm implements convolution directly instead of preprocessing with image to columns. Algorithm parameters are selected to meet constraints on global and shared memory access bandwidth, register usage, shared memory usage, and instructions per clock. Parallel arithmetic operations and memory access are achieved with several parallel blocks per streaming processor. Results comparing with state-of-the-art implementations are presented.
Fields of Study & Subjects
Advanced Neural Network Applications
Neural Networks and Applications
Adversarial Robustness in Machine Learning
Publication details
- Journal / source
- Cluster Computing
- Publisher
- Springer Science and Business Media LLC
- Volume
- 29
- Issue
- 2
- Publication date
- 9 January 2026
- Language
- EN