Module csr_queue_split 

Mixed queue traversal that keeps low-degree rows scalar and sends only hubs to row-strided teams. Mixed sparse CSR queue traversal for active sets with a small number of hubs.

A global row-strided pass is excellent for true hub rows, but wastes lanes on the many one-edge and three-edge rows that usually travel in the same active queue. This primitive keeps low-degree rows in a scalar queue pass and compacts only high-degree sources into a second queue for row-strided traversal.

Structs

CsrQueueSplitLowForwardCpuResult

CPU result for the low split pass.

CsrQueueSplitLowForwardParams

Positional inputs for csr_queue_split_low_forward_traverse.

Constants

CSR_QUEUE_SPLIT_HIGH_DEGREE_THRESHOLD

Degree at which a queued row has enough work to amortize a 32-lane team.

CSR_QUEUE_SPLIT_LOW_FORWARD_OP_ID

Canonical op id for mixed low-row traversal and high-row compaction.

CSR_QUEUE_SPLIT_LOW_FORWARD_WORKGROUP_SIZE

Workgroup shape for the low-row split pass.

Functions

csr_queue_split_low_dispatch_grid

Dispatch grid for the one-lane-per-active-source low split pass.

csr_queue_split_low_forward_traverse

Build the low-row half of a mixed queue traversal.

csr_queue_split_low_forward_traverse_with

Build the low-row half of a mixed queue traversal.

csr_queue_split_mixed_logical_lanes

Logical lanes consumed by low split plus a high row-strided follow-up pass.

try_csr_queue_split_low_forward_traverse_cpu

Fallible CPU reference for the low split pass.