Struct Buffer 

pub struct Buffer { /* private fields */ }

A device buffer that may be a view into another buffer.

Handle-identity (id) is per-Buffer value, including views — mirroring tinygrad, where each BUFFER_VIEW produces a distinct UOp identity. Storage-identity (the underlying Arc<BufferData>) is shared between a buffer and its views; use Buffer::storage_id to compare it.

Implementations

impl Buffer

pub fn new( allocator: Arc<dyn Allocator>, dtype: DType, shape: Vec<usize>, options: BufferOptions, ) -> Self

Create a new buffer with lazy allocation.

pub fn allocate( allocator: Arc<dyn Allocator>, dtype: DType, shape: Vec<usize>, options: BufferOptions, ) -> Result<Self>

Create a new buffer with immediate allocation.

pub fn view(&self, offset: usize, size: usize) -> Result<Self>

Create a view into this buffer.

The view shares storage with self (same Arc<BufferData>) but gets a fresh BufferId so the runtime parallel-hazard model treats disjoint views of one arena as independent — mirroring tinygrad’s BUFFER_VIEW-as-distinct-identity semantics. Use Buffer::storage_id to compare storage identity instead.

pub fn ensure_allocated(&self) -> Result<()>

Ensure the underlying buffer is allocated.

pub fn is_allocated(&self) -> bool

Check if the buffer is allocated.

pub fn size(&self) -> usize

Get the size of this buffer view in bytes.

pub fn offset(&self) -> usize

Get the offset of this view in bytes.

pub fn dtype(&self) -> DType

Get the data type.

pub fn shape(&self) -> &[usize]

Get the shape of this buffer.

pub fn as_host_bytes(&self) -> Result<&[u8]>

Get a byte slice of the buffer data (CPU-accessible buffers only).

Zero-copy. For realized tensors after realize(), this is safe because the scheduler guarantees no concurrent kernel writes.

Errors

• NotAllocated if buffer hasn’t been allocated
• NotCpuAccessible for CUDA device buffers (use copyout instead)

pub fn as_host_bytes_mut(&self) -> Result<&mut [u8]>

Get a mutable byte slice of the buffer data (CPU-accessible buffers only).

Safety contract (same as as_host_bytes)

Caller must ensure no kernels are executing concurrently.

Errors

• NotAllocated if buffer hasn’t been allocated
• NotCpuAccessible for CUDA device buffers

pub fn as_array<T: HasDType>(&self) -> Result<ArrayViewD<'_, T>>

Typed immutable view over CPU-accessible buffer memory.

Returns an ndarray view shaped according to the buffer’s concrete dimensions. Only works for CPU-accessible buffers — fails for device-only CUDA memory.

Errors

• TypeMismatch if T::DTYPE doesn’t match buffer dtype
• NotCpuAccessible for non-CPU-accessible buffers
• NotAllocated if buffer hasn’t been allocated

pub fn as_array_mut<T: HasDType>(&self) -> Result<ArrayViewMutD<'_, T>>

Typed mutable view over CPU-accessible buffer memory.

Same as Self::as_array but allows writes. Caller must ensure no kernels are executing concurrently (safety is the caller’s responsibility).

Errors

Same as Self::as_array.

pub fn as_slice<T: HasDType>(&self) -> Result<&[T]>

Zero-copy typed slice view (CPU-accessible only).

pub fn item<T: HasDType + Copy>(&self) -> Result<T>

Read a single scalar value from the buffer (CPU-accessible only).

Panics if the buffer contains more than one element.

pub fn allocator(&self) -> &dyn Allocator

Get the allocator used by this buffer.

pub fn allocator_arc(&self) -> Arc<dyn Allocator>

Get an Arc-cloned handle to the allocator, suitable for constructing new buffers on the same device (used by the arena memory planner to allocate per-lane arenas matching prototype buffers’ device).

pub fn id(&self) -> BufferId

Get the unique identifier for this buffer handle.

Each Buffer value (including each view) carries its own BufferId; disjoint views of one arena therefore have different ids. Used by the runtime parallel-hazard model. To compare storage identity (i.e. “do these two buffers share the same underlying allocation”), use Buffer::storage_id instead.

pub fn total_size(&self) -> usize

Size of the underlying allocation in bytes (shared by every view of this buffer’s storage). Distinct from Buffer::size, which returns the view’s size — for a non-view buffer the two are equal; for a view into an arena, total_size reports the arena’s allocation size while size reports just the view’s window.

pub fn storage_id(&self) -> BufferId

Stable identifier for this buffer’s underlying allocation.

Equal across a base buffer and all of its views, distinct between independent allocations. Unlike a heap-pointer probe, this id is minted once at allocation time and never reused — safe to use as a hash key or alias-detection key without worrying about allocator-reuse aliasing.

pub fn copyin(&mut self, src: &[u8]) -> Result<()>

Copy data from host memory into this buffer.

pub fn copyout(&self, dst: &mut [u8]) -> Result<()>

Copy data from this buffer to host memory.

pub fn copy_from(&mut self, src: &Buffer) -> Result<()>

Copy data from another buffer to this buffer.

pub fn synchronize(&self) -> Result<()>

Synchronize the device (wait for all operations to complete).

pub unsafe fn as_raw_ptr(&self) -> *mut u8

Get a raw pointer to the buffer data for kernel execution.

Safety

The returned pointer is only valid while the buffer is allocated. The caller must ensure:

• Buffer remains allocated during pointer lifetime
• No conflicting accesses occur during kernel execution
• Pointer is not used after buffer is freed

Panics

Panics if the buffer is not yet allocated.

Trait Implementations

impl Clone for Buffer

fn clone(&self) -> Buffer

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for Buffer

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

Formats the value using the given formatter.