Struct BlobHeader 

pub struct BlobHeader {

    pub size: u64,
    pub required_mask: u64,
    pub planes_bytes: u64,
    pub storage_kind: u32,
    pub pid: u32,
    pub fence_fd: i32,
    pub dtype: u32,
    pub quant_axis: i32,
    pub ndim: u32,
    pub plane_count: u32,
    pub quant_scales_len: u32,
    pub quant_zero_points_len: u32,
    pub strings_bytes: u32,
    pub strides_len: u32,
}

The blob’s fixed header. See the module docs for the byte layout.

Field order here mirrors the byte layout so the two can be read side by side, but this struct is never cast to or from bytes — see Self::write_to and parse_header, which go field by field in explicit little-endian.

Fields

size: u64

Total bytes of this blob, header included. The size prefix that makes the format append-only: an older consumer reads the prefix it knows and skips to size.

required_mask: u64

Bits describing additions that change interpretation rather than adding information. Any bit a consumer does not recognise means refuse.

planes_bytes: u64

Byte length of the planes region.

storage_kind: u32

Backing store for the whole tensor; the shared TensorMemory code.

pid: u32

Producer pid, for pidfd_open/pidfd_getfd and, on Windows, for OpenProcess(PROCESS_DUP_HANDLE). Zero when all planes are inlined, where there is no handle to reopen.

Trust boundary. This value and the handle values beside it come out of the blob, so a hostile one makes an importer probe a pid it chose and a handle slot in that process. That is the whole of what it buys: the import asks for PROCESS_DUP_HANDLE alone, and hands the duplicate only to OpenSharedResource1 / OpenSharedFence, which refuse anything that is not a shareable D3D11 object. Nothing here makes a blob safe to accept from an untrusted peer – import blobs from peers you would already hand a texture to.

fence_fd: i32

Acquire fence, or -1 for none. Per-tensor, never per-plane.

dtype: u32

Element type; the shared DType code.

quant_axis: i32

Per-channel quantization axis; -1 per-tensor, -2 none.

ndim: u32

Number of dimensions in the addressing grid.

plane_count: u32

Number of plane records in the planes region.

quant_scales_len: u32

Number of f32 quantization scales.

quant_zero_points_len: u32

Number of i32 zero points; zero means symmetric.

strings_bytes: u32

Byte length of the five-string region.

strides_len: u32

Number of stride entries: ndim, or 0 meaning densely packed C-order.

Carried rather than assumed — an empty stride array is a distinct value per Tensor.msg, not an absent one, and flattening the two would break the reference-mode round-trip identity.

Implementations

impl BlobHeader

pub fn empty() -> Self

An all-zero header, for tests and for .. update syntax.

pub fn write_to(&self, buf: &mut [u8])

Serialize into the first HEADER_LEN bytes of buf.

Panics

If buf is shorter than HEADER_LEN.

Trait Implementations

impl Clone for BlobHeader

fn clone(&self) -> BlobHeader

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for BlobHeader

impl Debug for BlobHeader

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

Formats the value using the given formatter.

impl Default for BlobHeader

fn default() -> BlobHeader

Returns the “default value” for a type.

impl Eq for BlobHeader

impl PartialEq for BlobHeader

fn eq(&self, other: &BlobHeader) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for BlobHeader