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

Mojo function

init_core_barriers

def init_core_barriers[num_input_stages: Int, num_accum_stages: Int](input_barriers_ptr: Pointer[SharedMemBarrier, MutUntrackedOrigin, address_space=AddressSpace.SHARED, _safe=False], input_consumer_count: Int32, accum_barriers_ptr: Pointer[SharedMemBarrier, MutUntrackedOrigin, address_space=AddressSpace.SHARED, _safe=False], accum_producer_arv_count: Int32, accum_consumer_arv_count: Int32, tmem_dealloc_ptr: Pointer[SharedMemBarrier, MutUntrackedOrigin, address_space=AddressSpace.SHARED, _safe=False], tmem_dealloc_thread_count: Int32)

Initialize input, output, and TMEM deallocation barriers.

Called inside the elect_one_warp && elect_one_thread guard. Handles the three barrier init steps shared by all SM100 kernels.

Parameters:

num_input_stages (Int): Number of input pipeline stages to initialize barriers for.
num_accum_stages (Int): Number of accumulator pipeline stages to initialize barriers for.

Args:

input_barriers_ptr (Pointer[SharedMemBarrier, MutUntrackedOrigin, address_space=AddressSpace.SHARED, _safe=False]): Shared-memory pointer to the input pipeline barriers.
input_consumer_count (Int32): Number of consumer arrivals on each input barrier.
accum_barriers_ptr (Pointer[SharedMemBarrier, MutUntrackedOrigin, address_space=AddressSpace.SHARED, _safe=False]): Shared-memory pointer to the accumulator pipeline barriers.
accum_producer_arv_count (Int32): Number of producer arrivals on each accumulator barrier.
accum_consumer_arv_count (Int32): Number of consumer arrivals on each accumulator barrier.
tmem_dealloc_ptr (Pointer[SharedMemBarrier, MutUntrackedOrigin, address_space=AddressSpace.SHARED, _safe=False]): Shared-memory pointer to the TMEM deallocation barrier.
tmem_dealloc_thread_count (Int32): Number of threads that arrive on the TMEM deallocation barrier.