OpenCL Caffe is an experimental deep neural network framework that ports Caffe to run on heterogeneous platforms using OpenCL, enabling computation across CPUs, GPUs, and APUs from various manufacturers.
The project addresses the need to accelerate deep learning workloads on diverse hardware by implementing OpenCL, an open standard parallel programming language. It ports the complete Caffe feature set to OpenCL, maintaining the same interfaces as the original framework so users can run existing models like AlexNet and VGG without modification. The implementation includes performance optimizations such as batched convolution operations using clBLAS, with configurable batch sizes to suit different hardware properties and input dimensions. The tool supports both OpenCL 2.0 and 1.2 standards.
Developers should be aware that this repository is explicitly marked as experimental and no longer under active development. The README recommends using the official OpenCL port of Caffe maintained in the BVLC repository instead. This project is most relevant for those working with AMD hardware, as it has been verified and tuned specifically on AMD devices including CPUs, GPUs, and APUs. The codebase provides three branches: stable for users, dev for contributors, and master synchronized with original Caffe.
Development on this project has ceased, with the README stating that no new development is happening. The repository was last actively maintained around August 2015 and represents an abandoned experimental effort that has been superseded by the official community port.
