Module occupancy_ext 

Extended occupancy helpers for CPU-side occupancy estimation.

Unlike the GPU-side queries in crate::occupancy that call cuOccupancy* driver functions, this module provides pure computation for analysing occupancy trade-offs without requiring a live GPU.

Features

• OccupancyCalculator — CPU-side occupancy estimation
• OccupancyGrid — sweep block sizes to find the optimum
• DynamicSmemOccupancy — occupancy with shared-memory callbacks
• ClusterOccupancy — Hopper+ thread block cluster support

Example

use oxicuda_driver::occupancy_ext::*;

let info = DeviceOccupancyInfo {
    sm_count: 84,
    max_threads_per_sm: 1536,
    max_blocks_per_sm: 16,
    max_registers_per_sm: 65536,
    max_shared_memory_per_sm: 102400,
    warp_size: 32,
};
let calc = OccupancyCalculator::new(info);
let est = calc.estimate_occupancy(256, 32, 0);
assert!(est.occupancy_ratio > 0.0);

Structs

ClusterConfig

Thread block cluster configuration for Hopper+ GPUs.

ClusterOccupancy

Hopper+ thread block cluster occupancy estimation.

ClusterOccupancyEstimate

Result of a cluster occupancy estimation.

DeviceOccupancyInfo

Hardware parameters needed for CPU-side occupancy estimation.

DynamicSmemOccupancy

Occupancy estimation where shared memory varies with block size.

OccupancyCalculator

CPU-side occupancy estimator — no GPU calls required.

OccupancyEstimate

Result of a CPU-side occupancy estimation for one configuration.

OccupancyGrid

Sweep block sizes to find the configuration that maximises occupancy.

OccupancyPoint

A single data point from a block-size sweep.

Enums

LimitingFactor

The resource that limits occupancy the most.