CCCL Cooperative

CCCL Cooperative is a set of device-side APIs for organizing groups of collaborating threads and executing collective algorithms. They are placed in the cuda::experimental::coop namespace and available through <cuda/experimental/coop/_header_> include files.

These APIs are successors to both CTK Cooperative Groups (<cooperative_groups.h>) and CUB thread-, warp-, and block-wide collective algorithms. The goal is to provide a more capable API that provides better flexibility, performance, and safety.

• CCCL Cooperative Groups
• CCCL Cooperative Algorithms

Warning

The CCCL Cooperative Groups implementation requires some experimental code to be enabled in libcu++ by defining _CUDAX_ENABLE_GROUP_FEATURES_IN_LIBCUDACXX before including any CCCL header.