Re: [eigen] Treat single-element Array as scalar for arithmetc operations

[Gael Guennebaud <gael.guennebaud@xxxxxxxxx>]

Hi,

what about writing a small helper function "as_scalar_or_array(x)" that would return a scalar if x is already a scalar or a 1x1 Eigen _expression_ at compile-time, and just x otherwise. This should solve most of your problems. Another direction would be to use "auto val = ArrayXd::Constant(n,x)" instead of "double val = x;" but this needs to know and store the size n.

Gaël.

On Thu, Jul 30, 2020 at 9:53 AM Mario Azcueta <mazcueta@xxxxxxxxx> wrote:

Thank you very much for your answers and suggestions.

Adrien's suggestion on overloading functions to accept both Arrays and scalars solved some of the problems (e.g. retrieving the size of an array/scalar)

However, I reached a wall when reductions were involved, i.e., there is a sub-function call that given a scalar returns an Array3d, and given a VectorXd returns an ArrayX3d. On these return values, I calculate the sum of each row, so that ArrayX3d reduces to an ArrayXd and the Vector3d reduces to a scalar. I'm still finding a way to deal with these 2 situations properly without replicating code. I may need to revise my approach since I brought it from Matlab, where it was very convenient, but I see it a bit forced given that Eigen does not allow summing single-element arrays with multiple-elements.

Thank you all again,

Mario

El mié., 29 jul. 2020 a las 9:15, Joseph Mirabel (<joseph.mirabel@xxxxxxx>) escribió:

Hi,

I have one comment on that. When I need this, I write code like

a[0] * b

where `a` is a vector of 1 element. This works fine but makes the code unclear as it is not evident that `a` has only one element. Of course, I can add an assert or a comment myself but that relies on me doing it...

I don't like the automatic casting but I would find an explicit method clearer. I am not able to find a good name for it. Maybe an explicit cast operation ?

a.asScalar() * b

a.singleElement() * b

static_cast<Scalar>(a) * b

Joseph

On 29/07/2020 03:56, Rasmus Munk Larsen wrote:

Generally speaking, silently treating vectors as scalars and vice versa with implied broadcasting is a common source of subtle bugs, so this behavior is unlikely to change. Eigen does support scalar*vector, if you extract the single element as such.

On Tue, Jul 28, 2020 at 6:06 AM Mario Azcueta <mazcueta@xxxxxxxxx> wrote:

Is there a way of adding/subtracting/multiplying two Eigen::ArrayXd arrays both when (a) they have dimension N, and (b) when one of them has dimension N and the other dimension 1?

I'd ideally like that the array with size=1 is treated like a scalar for these operations. I'm somewhat trying to replicate the way Matlab behaves.

As an example, I'd the code below to work both when a1 and a2 have the same size, as well as when one of them has size=1 and the other size=N:

ArrayXd add_arrays(const ArrayXd& a1, const ArrayXd& a2)
{
   return a1 + a2;
}

Of course, this could be solved with an if statement, but I want to avoid that since my actual code would require to write a long function verifying all combinations of sizes (they're many args). Also, writing a template function gets me problems when trying to use some ArrayXd methods and I pass a double.

I've looked into the EIGEN_ARRAYBASE_PLUGIN preprocessor which may seem like the way to extend this behavior, but I'm not sure this is the right way.. any advice will be welcome.

Note: I've also post this question here

https://stackoverflow.com/questions/63074198/c-eigen-treat-single-element-array-as-scalar-for-arithmetc-operations