Struct DotGeneralConfig 

pub struct DotGeneralConfig {
    pub lhs_contracting_dims: SmallVec<[usize; 4]>,
    pub rhs_contracting_dims: SmallVec<[usize; 4]>,
    pub lhs_batch_dims: SmallVec<[usize; 4]>,
    pub rhs_batch_dims: SmallVec<[usize; 4]>,
}

DotGeneral dimension configuration.

Each axis list stores up to four axes inline and spills to the heap for larger contractions; this is not a tensor-rank limit. Records only the dim-numbering roles (contracting / batch; free is derived). Rank info travels with the enclosing StdTensorOp::DotGeneral variant at the trace/StdTensorOp layer, and with ExecInstruction::output_shapes at the exec layer. This separation makes it structurally impossible for stored ranks to drift from actual tensor ranks (issue #664).

Output layout

The output shape is [lhs_free..., rhs_free..., batch...] (col-major batch-trailing convention): batch axes come last, unlike PyTorch’s batch-leading bmm. Batch dims have the largest stride so that each batch slice occupies a contiguous block of memory. Free axes keep their input order. For attention scores, q[d, lq, b] against k[d, lk, b] with d contracted and b batched gives [lq, lk, b]; see the “Contraction Output Layout” section of the tensor-operations guide for a worked example. Transpose afterwards only where a consumer needs batch first.

Examples

use tenferro_tensor::DotGeneralConfig;

let config = DotGeneralConfig {
    lhs_contracting_dims: [1].as_slice().into(),
    rhs_contracting_dims: [0].as_slice().into(),
    lhs_batch_dims: [].as_slice().into(),
    rhs_batch_dims: [].as_slice().into(),
};

Fields

lhs_contracting_dims: SmallVec<[usize; 4]>

rhs_contracting_dims: SmallVec<[usize; 4]>

lhs_batch_dims: SmallVec<[usize; 4]>

rhs_batch_dims: SmallVec<[usize; 4]>

Implementations

impl DotGeneralConfig

pub fn validate_dims_with_ranks( &self, lhs_rank: usize, rhs_rank: usize, ) -> Result<()>

Validate that all dimension indices are within range for the given explicit ranks and that no axis appears in multiple roles.

Call sites supply the actual operand ranks (from the tensor shapes they have in hand). The config itself carries only the dim-numbering roles.

Examples

use tenferro_tensor::DotGeneralConfig;

let config = DotGeneralConfig {
    lhs_contracting_dims: [1].as_slice().into(),
    rhs_contracting_dims: [0].as_slice().into(),
    lhs_batch_dims: [].as_slice().into(),
    rhs_batch_dims: [].as_slice().into(),
};
config.validate_dims_with_ranks(2, 2).unwrap();

Errors

Returns crate::Error::Validation with an axis, duplicate-axis, or configuration source when the dimension roles are invalid.

Trait Implementations

impl Clone for DotGeneralConfig

fn clone(&self) -> DotGeneralConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for DotGeneralConfig

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

Formats the value using the given formatter.

impl Eq for DotGeneralConfig

impl Hash for DotGeneralConfig

fn hash<__H: Hasher>(&self, state: &mut __H)

Feeds this value into the given Hasher.

fn hash_slice<H>(data: &[Self], state: &mut H)

where H: Hasher, Self: Sized,

Feeds a slice of this type into the given Hasher.

impl PartialEq for DotGeneralConfig

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

Equality operator ==.

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

Inequality operator !=.

impl StructuralPartialEq for DotGeneralConfig