Struct GlobalMemoryPool 

pub struct GlobalMemoryPool { /* private fields */ }

Enhanced global memory pool with SciRS2 memory optimization

Implementations

impl GlobalMemoryPool

pub fn new() -> Self

Create a new enhanced global memory pool with SciRS2 integration

pub fn create_large_tensor<T>( &mut self, shape: &[usize], device: DeviceType, ) -> Result<Tensor<T>>

where T: Clone + Default + TensorElement,

✅ SciRS2 Memory-Efficient Tensor Creation for Large Tensors

pub fn create_lazy_tensor<T>( &mut self, shape: &[usize], device: DeviceType, ) -> Result<Tensor<T>>

where T: Clone + Default + TensorElement,

✅ SciRS2 Lazy Tensor Creation - Defer allocation until needed

pub fn create_zero_copy_view<T>( &self, source: &Tensor<T>, offset: usize, shape: &[usize], ) -> Result<Tensor<T>>

where T: Clone + TensorElement,

✅ SciRS2 Zero-Copy Operations for efficient tensor views

pub fn get_enhanced_stats(&self) -> PoolStatistics

Get memory usage statistics enhanced with SciRS2 metrics

pub fn acquire_uninit<T: 'static>(&mut self, count: usize) -> ReusedBuffer<T>

Acquire a truly-pooled, uninitialized buffer for count elements of type T with natural alignment (align_of::<T>()).

This is the low-level method. Prefer the free function global_acquire_uninit.

The returned ReusedBuffer<T> holds the actual pooled allocation — no copy is made. Callers must initialize all elements before reading them.

For custom alignment (e.g. 32-byte SIMD alignment), use Self::acquire_uninit_aligned instead.

pub fn acquire_uninit_aligned<T: 'static>( &mut self, count: usize, align: usize, ) -> ReusedBuffer<T>

Acquire a pooled uninitialized buffer with custom alignment.

Useful for SIMD-aligned buffers (32-byte for AVX2, 64-byte for AVX-512, etc.). Buffers acquired with a given align go into their own bucket keyed by (TypeId, SizeClass, align), so they never collide with naturally-aligned allocations of the same type+size.

Panics

• if align is not a power of two
• if align < std::mem::align_of::<T>()

pub fn allocate<T: TensorElement + Default + 'static>( &mut self, count: usize, ) -> Vec<T>

👎Deprecated:

Use global_acquire_uninit instead for zero-copy buffer reuse

Allocate memory for tensor elements.

Returns a zero-initialized Vec<T>.

Deprecation

Use global_acquire_uninit for zero-copy buffer reuse.

pub fn find_size_class(&self, size_bytes: usize) -> usize

Find appropriate size class for allocation

pub fn deallocate<T: 'static>(&mut self, data: Vec<T>)

Deallocate memory by dropping it (legacy; buffer is not returned to pool).

The deallocate method previously attempted to store the allocation in the pool using an unsafe Vec<u8> transmutation that could not reconstruct the correct layout. Now the Vec is simply dropped. Use ReusedBuffer::release_to_pool for true pool return.

pub fn clear(&mut self)

Clear all pools

pub fn get_statistics(&self) -> &PoolStatistics

Get basic statistics

pub fn hit_rate(&self) -> f64

Get cache hit rate

pub fn cleanup(&mut self)

Cleanup unused memory

Trait Implementations

impl Debug for GlobalMemoryPool

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for GlobalMemoryPool

fn default() -> Self

Returns the “default value” for a type.