Struct TensorDesc 

#[repr(C)]
pub struct TensorDesc {

    pub version: u32,
    pub kind: u32,
    pub handle: i64,
    pub ptr: SendPtr,
    pub ndim: u32,
    pub dtype: u32,
    pub shape: [u64; 8],
    pub strides: [i64; 8],
    pub fourcc: u32,
    pub format: u32,
    pub colorimetry: u32,
    pub flags: u32,
    pub capacity: u64,
    pub sync: u64,
}

The cross-package descriptor. Layout is part of the contract: it is #[repr(C)] and its size is pinned by test.

Fields

version: u32

ABI_VERSION. Checked, never assumed.

kind: u32

One of the kind constants.

handle: i64

dma-buf fd, IOSurface id or GL buffer id; -1 when unused.

ptr: SendPtr

Host pointer when addressable, else null – except under kind::PBO, where it instead carries a *const crate::pbo::PboOpsVtable, or null if the producer attached none. Interpretation is keyed by Self::kind, the same established pattern Self::sync’s own doc comment already documents for that field: a PBO tensor never takes a HostPin (PboTensor refuses pin_host — the address cannot outlive the map), so this field is otherwise permanently unused for that kind, which is exactly why it is the field that carries the vtable instead of a new one requiring a layout change. Valid under the exact same borrow contract as the host-pointer case: only while the producer’s own keepalive (for the capsule protocol, the capsule’s) keeps the owning tensor alive.

ndim: u32

Number of significant entries in shape and strides.

dtype: u32

Element type, one of the [dtype] constants.

Deliberately placed here: ndim leaves a 4-byte hole before the 8-aligned shape, so this field costs nothing. Without it strides — documented in elements — cannot be converted to bytes by a consumer, and a Tensor<f32> is indistinguishable from a Tensor<u8> of the same shape.

shape: [u64; 8]

Logical shape, most-significant dimension first.

strides: [i64; 8]

Strides in bytes, signed so a consumer can express flips.

Bytes rather than elements so a hardware row pitch is always representable: a pitch need not be a whole number of dtype elements (and a sub-byte dtype has no element size to divide by). Multiply a coordinate by the stride directly; do not scale by dtype size.

fourcc: u32

DRM FourCC of the pixel format, or 0 for a non-image tensor. Kept for third-party/DRM interop; a HAL-aware consumer should prefer Self::format and fall back to this only when format is 0 (see format’s doc for why fourcc alone is insufficient).

format: u32

One of the format constants, or 0 for a non-image tensor.

colorimetry: u32

Packed colorimetry, or 0 when unspecified.

flags: u32

Bit flags; see flags. Zero when nothing applies.

Occupies what was a 4-byte alignment hole before capacity, so it cost nothing to add.

capacity: u64

Bytes available in the producer’s underlying allocation, from this tensor’s window start (its TensorTrait::capacity_bytes ) – may exceed the byte size shape/strides imply when the producer over-allocated: a pool tensor sized for its largest expected image but holding a smaller one today, or row-pitch/MCU-alignment padding a decoder needs beyond the logical width. Without this, a consumer importing a HOST descriptor has no way to learn that headroom exists and must treat the declared shape as the allocation’s exact size – which starves a later configure_image/set_logical_shape of capacity the producer actually has. DMABUF/IOSURFACE consumers query their real capacity from the handle itself and do not depend on this field, but it is populated for every kind for consistency.

sync: u64

Completion handle for work still in flight against this tensor, or 0 when flags::SYNC_PRESENT is clear.

Reserved: producers in this build always leave it clear. It is declared now because the field has to exist before the first release or adding it later costs a capsule-name bump, and it is the one gap the synchronous-at-the-boundary contract cannot close: a foreign producer with queued GPU work has no other way to say “wait on this first”.

Interpretation is keyed by Self::kind, since each backing already implies its own fence flavour and a separate discriminant would be redundant:

• kind::DMABUF – a sync_file fd from DMA_BUF_IOCTL_EXPORT_SYNC_FILE. The consumer owns it and must close it.
• kind::PBO – a GLsync from glFenceSync, valid only in the producer’s share group.
• kind::CUDA_DEVICE – a cudaEvent_t.
• kind::HOST, kind::IOSURFACE – no fence flavour defined; SYNC_PRESENT must be clear.

Implementations

impl TensorDesc

pub const MAX_NDIM: usize = 8

Maximum rank the fixed-size shape/strides arrays can carry.

pub fn shape(&self) -> &[u64]

The logical shape as a slice.

pub fn strides(&self) -> &[i64]

The strides as a slice, in bytes.

pub fn len(&self) -> u64

Total element count.

pub fn is_empty(&self) -> bool

True when the descriptor addresses no elements.

Trait Implementations

impl Clone for TensorDesc

fn clone(&self) -> TensorDesc

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for TensorDesc

impl Debug for TensorDesc

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

Formats the value using the given formatter.