bmm_op

→TensorImplPtr

int bmm_op(const int & a, const int & b)

source edit

Python wrappers

pylucid.ops.bmmfunction

Strict 3-D batched matrix multiply: $C_b = A_b \cdot B_b$ .

Composite — rank-checked alias of matmul_op. Rejects inputs of rank $\ne 3$ and demands matching leading batch dimensions so the error path is crisp. Gradient flows through MatmulBackward.

Math

For every batch index $b \in [0, B)$ :

C_{b, i, j} = \sum_{k} A_{b, i, k} \, B_{b, k, j}

See Also

mm_op, matmul_op.

Parameters

aTensorImplPtr

Left operand of shape (B, M, K).

bTensorImplPtr

Right operand of shape (B, K, N). Leading batch B must match a.

Returns

TensorImplPtr

Tensor of shape (B, M, N).

Raises

Failure

If either input is null, not exactly 3-D, or batch dimensions disagree.