Struct KernelScheduler 

pub struct KernelScheduler { /* private fields */ }

Kernel dependency graph and launch-order scheduler.

Kernels are registered via add_kernel and dependencies added via add_dependency. Once the graph is complete, launch_order returns a topological ordering that satisfies all constraints.

Implementations

impl KernelScheduler

pub fn new() -> Self

Create an empty scheduler.

pub fn add_kernel(&mut self, spec: KernelSpec) -> Result<(), SchedulerError>

Register a kernel with the scheduler.

Errors

Returns SchedulerError::DuplicateKernel if a kernel with the same ID has already been registered.

pub fn add_dependency( &mut self, dependent: u32, dependency: u32, ) -> Result<(), SchedulerError>

Declare that kernel dependent must not start until kernel dependency has finished.

Errors

• SchedulerError::KernelNotFound if either ID is not registered.
• SchedulerError::CyclicDependency if the edge would introduce a cycle.

pub fn dependencies_of( &self, kernel_id: u32, ) -> Result<Vec<u32>, SchedulerError>

Return the IDs of all direct dependencies of kernel_id.

Errors

Returns SchedulerError::KernelNotFound if the ID is not registered.

pub fn launch_order(&self) -> Result<Vec<u32>, SchedulerError>

Compute a valid topological launch order for all registered kernels.

Uses Kahn’s algorithm with a min-heap (via BTreeSet) for deterministic output: among ready kernels, the one with the smallest ID is picked first.

Errors

Returns SchedulerError::CycleDetected if the graph contains a cycle (which should not happen if add_dependency correctly enforces the acyclicity invariant, but is checked defensively here).

pub fn occupancy( &self, kernel_id: u32, sm_warp_limit: u32, warp_size: u32, ) -> Result<OccupancyEstimate, SchedulerError>

Compute occupancy for a specific kernel.

Errors

Returns SchedulerError::KernelNotFound if the ID is not registered.

pub fn simulate_warp_stats( &self, sm_warp_limit: u32, warp_size: u32, ) -> Result<Vec<WarpStats>, SchedulerError>

Simulate execution and return warp statistics for each kernel in launch order.

The simulation model:

• Active warps = min(warps_per_group * work_groups, sm_warp_limit).
• Stalled warps = max(0, total_warps_launched − active_warps).

Errors

Returns an error if a valid launch order cannot be produced.

pub fn kernel_count(&self) -> usize

Number of kernels registered in the scheduler.

pub fn spec(&self, kernel_id: u32) -> Option<&KernelSpec>

Retrieve the KernelSpec for a given ID, if registered.

Trait Implementations

impl Default for KernelScheduler

fn default() -> Self

Returns the “default value” for a type.