@grid_barriers

In a @persistent kernel the compiler places a grid barrier wherever one block may read or write device memory another block wrote since the last one, between two of the kernel’s top-level statements and never inside a loop that hands out units. Each barrier costs every block a wait for the slowest.

@grid_barriers(n)

@grid_barriers(n) asserts the kernel’s total, the barriers it writes and the ones the compiler places, and fails compilation with both counts when they differ. A change that adds a barrier to a chain then fails where the chain is compiled, rather than costing time quietly. PHOBOS_LOG=debug names each placed barrier with the two accesses it orders. See the Attributes table in the spec.