Compare occupancy-based LRB grids with the fixed SM multiplier #12472

Area

CUB

Problem

The warp and small-block sorting entries currently use an SM-scaled grid multiplier of 128. A grid based on each kernel's achievable occupancy may reduce idle work or improve scheduling, but a single resident wave can also underutilize irregular workloads.

Proposed work

• Compare the existing multiplier with occupancy-based grids for warp and small-block entries independently.
• Use each compiled kernel's actual threads, registers, and shared-memory requirements; evaluate a tunable number of occupancy-sized waves rather than equating occupancy with total work.
• Preserve device-resident work bounds and avoid a device-to-host count copy for grid sizing.
• Keep the grid multiplier distinct from the launch-bounds residency target and include the cost of resource queries in ordinary execution.

Validation / completion criteria

• Benchmark uniform, skewed, few-segment, and very large workloads across supported key types and architectures.
• Report end-to-end and per-stage time, idle work, achieved occupancy, and ordinary versus captured execution.
• Retain the fixed-grid option unless measured results support a replacement heuristic; rerun correctness/synchronization checks when grids undersubscribe or oversubscribe work.

Context

Follow-up to the LRB segmented-sort backend in #12236.

Related work: #10952, #10800.