Question about add operation count in different case

[LOSKIw]

After going through the source code, I found that in different func the addition operation is count in different style

in torch.add all addition seems to be counted

while in some matrices multiplication, part of the addition is not

one example is in torch,addmm,

input + mat1 @ mat2

def _addmm_tensor_flops_hook(input, mat1, mat2, *, beta=1, alpha=1, out=None):
    flops = np.prod(mat1.shape, dtype=np.int64) * mat2.shape[-1]
    if beta != 0:
        flops += np.prod(input.shape, dtype=np.int64)
    return flops

the addition of input + new matrix is included, while the addition in mat1 @ mat2 seems not been considered

assume mat1 and mat2 have shape m*n and n*l, the total operation output for addmm will be m*n*l+m*l, but I think it might be m*n*l+(n-1)*m*l+m*l=2*m*n*l

I understand that addition might affect little to the total process, but maybe a constant regulation for computing will be more reasonable? Or is there any error in my expression, like in real case we could have some extra condition?

Thank you very much