For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Mojo function

syncwarp

def syncwarp(mask: Int = Int(-1))

Synchronizes threads within a warp using a barrier.

This function creates a synchronization point where threads in a warp must wait until all threads specified by the mask reach this point. On NVIDIA GPUs, it uses warp-level synchronization primitives. On AMD GPUs, this acts as a wave execution barrier. On Apple GPUs, this acts as a SIMDGROUP execution barrier. Lane masks are not supported, so the mask argument is ignored and all active lanes must reach this point.

Note:

• On NVIDIA GPUs, this maps to the nvvm.bar.warp.sync intrinsic.
• On AMD GPUs, this maps to the llvm.amdgcn.wave.barrier intrinsic.
• On Apple GPUs, this provides execution synchronization only via a SIMDGROUP barrier with mem_none (no memory fence). Use barrier() for threadgroup memory ordering.
• Threads not participating in the sync must still execute the instruction.

Args:

• ​mask (Int): An integer bitmask specifying which lanes (threads) in the warp should be synchronized. Each bit corresponds to a lane, with bit i controlling lane i. A value of 1 means the lane participates in the sync, 0 means it does not. Default value of -1 (all bits set) synchronizes all lanes.