Struct DeviceArena 

pub struct DeviceArena { /* private fields */ }

Power-of-two bucketed free list of f64 device slices.

Allocations round the requested element count up to the next usize::next_power_of_two. On drop the slab is handed back to the arena under the same bucket via DeviceArena::release. Held under a Mutex by every backend that uses it because large-scale fits dispatch from multiple rayon workers; the mutex is only held during alloc / release, never across kernel launches.

Implementations

impl DeviceArena

pub fn bucket_of(elements: usize) -> usize

pub fn alloc( &mut self, stream: &Arc<CudaStream>, elements: usize, label: &'static str, ) -> Result<(usize, CudaSlice<f64>), GpuError>

Allocate a device slice of at least elements f64s. Returns the bucket size actually allocated so the caller can release into the same bucket on drop. label is woven into the error message if the underlying alloc_zeros fails so failures stay attributable to the originating backend (matching the pre-extraction wording).

pub fn release(&mut self, bucket: usize, slab: CudaSlice<f64>)

Trait Implementations

impl Default for DeviceArena

fn default() -> DeviceArena

Returns the “default value” for a type.