Cooperative Groups

Raw masks and intrinsics work, but they are fiddly. Cooperative groups wrap threads into objects you can name, size, and synchronize explicitly.

Start by getting a handle to your block of threads. this_thread_block returns a group you can sync just like syncthreads, but as an object.

Calling sync on the block is the modern barrier. It does exactly what syncthreads does, but reads clearly as a method on the group you mean.

You can split a block into fixed-size tiles. A 32-lane tile gives you a warp-sized group with clean methods instead of raw shuffle masks.

A tiled group offers shfl_down and friends without a mask argument. The group already knows its members, so the API stays short and safe.

Every group exposes its members and your spot. thread_rank gives your index inside the group, and size returns how many threads it holds.

Tiles need not be 32 wide. A tiled_partition of 8 or 16 makes sub-warp groups, handy when your data naturally clusters in small sets.

The library ships ready-made collectives. A group reduce sums a tile in one call, hiding the offset loop you wrote by hand earlier.

The boldest group is the grid group. With a cooperative launch, every block can sync at one barrier, something a normal kernel cannot do.

Grid-wide sync only works if you start the kernel with cudaLaunchCooperativeKernel and the GPU supports it. A normal launch will not allow it.

Cooperative groups make warp code readable and portable: no hand-managed masks, clear scopes, and reusable collectives that match what you mean.

Recall how you create a warp-sized cooperative group from a block.

Cooperative groups turn masks into named objects: tile a block, call reduce, even sync a whole grid. You now own warp-level CUDA. ✨