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
wait_on_producers
def wait_on_producers()
Waits on the predecessor grids, then synchronizes the CTA.
The CTA barrier is a defence, not a measured requirement. An earlier layout
of the Mamba-2 kernels showed the assembler issuing LDG.E.CONSTANT loads
(what a tensor with an immutable origin loads with) above
griddepcontrol.wait once their address was known, so they read what the
predecessor had not written yet. A CTA barrier is an ordering point those
loads are not moved across. In the current kernels (the conv decode launch,
the decode and staged scan tiles, both norms) the SASS has no global load
above the wait with or without the barrier, and no test exercises the
barrier. That holds for the tile in use: re-check the SASS when a kernel's
tile or load order changes. Call this where every thread of the CTA arrives,
before any early return.