There are 3 Low-level APIs for Nvidia GPGPU Computing

• CUDA Runtime API
• CUDA Driver API
• OpenCL

OpenCL offers a general API that is supported by many vendors and allows the potential to run kernels efficiently on the CPU in addition to coprocessor devices. CUDA Driver is a more mature GPGPU programming API with stable compilers, freedom in the choice of host compilers, and can potentially generate the most efficient code for Nvidia devices. CUDA Runtime offers a more succinct API and support for GPU code integrated with host code.

Unfortunately, all 3 APIs have different function names and data types. Memory management is complicated:

• Host, Host Pinned, Host Pinned Write-Combined, Host Portable, Device, Device 2D with Padding, Constant Device, Texture 2D, Texture 3D, etc.
• CUDA-RT has 20 memory copy functions requiring up to 8 parameters
• CUDA-Driver has 17 memory copy functions
• OpenCL has 8 memory copy functions requiring up to 11 parameters

Geryon

Geryon is intended to be a simple library for managing all three APIs with a consistent interface

• Change from one API to another by simply changing the namespace
• Use multiple APIs in the same code
• Lightweight (only include files – no build required)
• Manage device query and selection
• Simple vector and matrix containers
• Simple routines for data copy and type casting
• Simple routines for data I/O
• Simple classes for managing device timing
• Simple classes for managing kernel compilation and execution

Top of page