Struct BFloat16Type 

pub struct BFloat16Type {}

The bfloat16 data type

This implements the ArrowFloatType trait for bfloat16 values.

Trait Implementations

impl ArrowFloatType for BFloat16Type

const FLOAT_TYPE: FloatType = FloatType::BFloat16

const MIN: Self::Native = bf16::MIN

const MAX: Self::Native = bf16::MAX

type Native = bf16

type ArrayType = FixedSizeBinaryArray

Arrow Float Array Type.

fn empty_array() -> Self::ArrayType

Returns empty array of this type.

impl Debug for BFloat16Type

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

Formats the value using the given formatter.

impl FloatArray<BFloat16Type> for FixedSizeBinaryArray

fn as_slice(&self) -> &[bf16]

Returns the underlying bf16 values as a borrowed slice.

Preconditions

• value_length() must be 2 (the FixedSizeBinary(2) storage shape used by BFloat16Array). Asserted at entry.
• The value buffer must be at least 2-byte aligned. Lance’s in-tree constructors always satisfy this: value buffers are built either via MutableBuffer (aligned to arrow-buffer’s ALIGNMENT constant, ≥32 bytes) or via Buffer::from_vec::<u16> (aligned to align_of::<u16>() == 2); both meet bf16’s 2-byte requirement. Externally-built FixedSizeBinaryArrays arriving via FFI, IPC, or Buffer::from_custom_allocation are not required by arrow-rs to be aligned beyond a single byte; passing one to this method violates the precondition. A debug_assert below catches such inputs in debug and test builds.

Endianness

lance-arrow is gated on target_endian = "little" at the crate root, so this method always returns values in the same byte order Lance writes (see BFloat16Array::value and the FromIterator impls).

type FloatType = BFloat16Type

fn from_values(values: Vec<bf16>) -> Self

Construct an array from a vector of values.

fn from_iter_values(values: impl IntoIterator<Item = T::Native>) -> Self

where Self: Sized,

Construct an array from an iterator of values.