MGPU¶

Introduction¶

The MGPU library strives to simplify the implementation of high performance applications and algorithms on multi-GPU systems. Its main goal is both to abstract platform dependent functions and vendor specific APIs, as well as simplifying communication between different compute elements. The library is currently an alpha release containing only limited yet already useful functionality such as

• segmented device vector (one vector spread across different GPUs)
• communication functions (simple and adaptable copy and broadcast)
• methods to dispatch jobs to GPUs
• multi-GPU management and synchronization
• basic support for FFT and BLAS functions
• vendor-specific API abstraction (currently only CUDA support)

The library is currently tested on Ubuntu 10.04.3 LTS with gcc 4.4.3 and compute capability 2.0 CUDA devices.

Things to expect in the near future:

• more complete documentation/reference
• more tested platforms, Windows support
• more complete FFT/BLAS support
• reduce/reduce_all communication functions