Struct MatrixBlockSlice 

pub struct MatrixBlockSlice<'a> {
    limbs: *const u64,
    stride: NonZeroUsize,
    _marker: PhantomData<&'a ()>,
}

A non-contiguous view of a 64 x 64 block within a larger matrix.

The block is stored in row-major order with a configurable stride between rows. This allows efficient access to sub-blocks within a matrix without copying data.

Safety

The limbs pointer must remain valid for the lifetime 'a, and must point to at least 64 valid rows spaced stride limbs apart.

Fields

limbs: *const u64

stride: NonZeroUsize

Number of limbs between consecutive rows

_marker: PhantomData<&'a ()>

Implementations

impl<'a> MatrixBlockSlice<'a>

pub(super) fn new(limbs: *const u64, stride: NonZeroUsize) -> Self

pub(crate) fn limbs(&self) -> *const u64

pub(crate) fn stride(&self) -> NonZeroUsize

pub fn iter(self) -> impl Iterator<Item = &'a u64>

Returns an iterator over the 64 rows of this block.

Safety

Each element is obtained via self.limbs.add(i * self.stride), which is safe because the constructor guarantees 64 valid rows at the given stride.

pub fn gather(self) -> MatrixBlock

Gathers the non-contiguous block into a contiguous MatrixBlock.

This operation is necessary before performing block-level GEMM since the AVX-512 kernel expects contiguous data.

Trait Implementations

impl Send for MatrixBlockSlice<'_>

impl Sync for MatrixBlockSlice<'_>