Overview

CULIB is an open-source library providing fast and flexible software for every layer of the CUDA programming model. The general goal of this project is to extend algorithms beyond warp, block, and device levels. CULIB also implements multi-GPU and multi-node versions of its algorithms.

CULIB is GPU oriented. It means that a result of any algorithm is stored on GPU and never copied on CPU. This also means that any CULIB components can be used only within kernel or device functions.

CULIB is not oriented on rapid prototyping of CUDA applications. Instead, CULIB is designed to raise the abstraction of complex algorithms without performance overhead. CULIB relies on policy-based design to let you tune the algorithms for specific cases.

Examples

CULIB provides algorithms as templated classes. Objects of these classes could be reused. The following code illustrates a kernel fragment performing exclusive scan across the threads of a warp:

using namespace culib;
warp::scan<int> scan;

const int sum_result = scan.exclusive (flag);
const int max_result = scan.exclusive (in[sum_result], binary_op::max<data_type> {});