matmul_op

→TensorImplPtr

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

Compute $A B$ with full NumPy-style batch broadcasting and autograd.

Thin public wrapper that delegates to MatmulBackward::forward.

Math

C[\ldots, i, j] = \sum_{k=0}^{K-1} A[\ldots, i, k]\, B[\ldots, k, j]

See Also

dot_op : Rank-1 / rank-2 specialised dispatch. MatmulBackward : Autograd node implementing the backward pass.

aTensorImplPtr

Left operand, rank

\ge 2

bTensorImplPtr

Right operand, rank

\ge 2

TensorImplPtr

Result tensor $C = A B$ with broadcast batch shape and trailing $(M, N)$ .

FLOPs per output element ≈ $2K$ ; total $\approx 2 \cdot B \cdot M \cdot N \cdot K$ where $B$ is the broadcast batch size.

    // C++: 2-D × 2-D plain GEMM.
    auto C = matmul_op(A, B);  // A: [M,K], B: [K,N]  → C: [M,N]