Struct CudaMemPool 

pub struct CudaMemPool { /* private fields */ }

A CUDA memory caching allocator.

Maintains per-dtype free lists keyed by element count. Reuses buffers when possible, falling back to cudaMalloc on cache miss.

Implementations

impl CudaMemPool

pub fn new() -> Self

Create a new empty memory pool.

pub fn alloc_f32( &self, dev: &Arc<CudaDevice>, n: usize, ) -> Result<CudaSlice<f32>, DriverError>

Allocate n elements of type f32, reusing a cached buffer if available. The returned buffer content is undefined (not zeroed).

pub fn alloc_zeros_f32( &self, dev: &Arc<CudaDevice>, n: usize, ) -> Result<CudaSlice<f32>, DriverError>

Allocate n elements of f32 and zero them.

pub fn alloc_f64( &self, dev: &Arc<CudaDevice>, n: usize, ) -> Result<CudaSlice<f64>, DriverError>

Allocate n elements of type f64.

pub fn alloc_zeros_f64( &self, dev: &Arc<CudaDevice>, n: usize, ) -> Result<CudaSlice<f64>, DriverError>

pub fn alloc_u16( &self, dev: &Arc<CudaDevice>, n: usize, ) -> Result<CudaSlice<u16>, DriverError>

Allocate n elements of type u16 (used for F16/BF16 storage).

pub fn alloc_zeros_u16( &self, dev: &Arc<CudaDevice>, n: usize, ) -> Result<CudaSlice<u16>, DriverError>

pub fn alloc_u8( &self, dev: &Arc<CudaDevice>, n: usize, ) -> Result<CudaSlice<u8>, DriverError>

Allocate n elements of type u8.

pub fn alloc_zeros_u8( &self, dev: &Arc<CudaDevice>, n: usize, ) -> Result<CudaSlice<u8>, DriverError>

pub fn alloc_u32( &self, dev: &Arc<CudaDevice>, n: usize, ) -> Result<CudaSlice<u32>, DriverError>

Allocate n elements of type u32.

pub fn alloc_zeros_u32( &self, dev: &Arc<CudaDevice>, n: usize, ) -> Result<CudaSlice<u32>, DriverError>

pub fn alloc_i64( &self, dev: &Arc<CudaDevice>, n: usize, ) -> Result<CudaSlice<i64>, DriverError>

Allocate n elements of type i64.

pub fn alloc_zeros_i64( &self, dev: &Arc<CudaDevice>, n: usize, ) -> Result<CudaSlice<i64>, DriverError>

pub fn reclaim_f32(&self, s: CudaSlice<f32>)

pub fn reclaim_f64(&self, s: CudaSlice<f64>)

pub fn reclaim_u16(&self, s: CudaSlice<u16>)

pub fn reclaim_u8(&self, s: CudaSlice<u8>)

pub fn reclaim_u32(&self, s: CudaSlice<u32>)

pub fn reclaim_i64(&self, s: CudaSlice<i64>)

pub fn reclaim_storage(&self, storage: CudaStorage)

Reclaim all buffers inside a CudaStorage, returning them to the pool.

pub fn empty_cache(&self)

Release all cached buffers back to the CUDA driver. This actually frees GPU memory.

pub fn stats(&self) -> PoolStats

Return a snapshot of pool statistics.

pub fn reset_stats(&self)

Reset hit/miss counters.

Trait Implementations

impl Default for CudaMemPool

fn default() -> Self

Returns the “default value” for a type.

impl Send for CudaMemPool

impl Sync for CudaMemPool