Broadcasting?

[xloem]

I'm unsure how to broadcast assignment to across columns or rows.

The wiki states:

// assign a vector to all the columns of a 3D array
A(0,0,all) = vec(all);

But this looks like it selects the first column and first row (index 0,0): what is the actual situation here?

[romeric]

Are you comparing Fastor with MATLAB/Fortran/Julia style broadcasting?

Note that Fastor tensors are row-major (C-style or NumPy style) so the order is swapped in that:

Tensor<double,3> vec = {1,2,3};      // a vector
Tensor<double,2,4,3> A(0);             // 3D tensor of all zeros 
A(0,0,all) = vec(all);
print(A);

will print

[0,:,:]
[1, 2, 3]
[0, 0, 0]
[0, 0, 0]
[0, 0, 0]
[1,:,:]
[0, 0, 0]
[0, 0, 0]
[0, 0, 0]
[0, 0, 0]

Here A is a 3D tensor made up of two 4x3 matrices.

So in A(i,j,k), i will give you the ith 4x3 matrix, j will give you the jth row in that matrix and k will give you the kth column in that matrix.

So in A(0,0,all) you are asking Fastor to give you the first 4x3 matrix and from that matrix give you the first row and all the columns.

Hope that is clear.

[xloem]

I see ... so by "assign a vector to all the columns" you mean that each column gets one different element of the vector.
How do we assign the same vector to every full column? such that all rows are filled with the same value, different for each row, and all columns become the same new vector?

[romeric]

Right I see that the documentation is incorrect. I will correct that. We don't have much documentation for Fastor at the moment.

What you are asking is essentially tiling numpy.tile, numpy.repeat or MATLAB's repmat feature to broadcast a vector to all elements of a tensor. This feature is unfortunately not implemented yet, but there is a high request rate for it. I will implement this as soon as I get some time.

Something like

Tensor<double,1,3> vec = {{1,2,3}};
Tensor<double,3,3> A = repeat<1,3>(vec); // repeat 3 times, column-wise (1 for column-wise and 3 for number of repetition)

[1, 1, 1]
[2, 2, 2]
[3, 3, 3]

For the time being however, unfortunately you will have to implement this with for loops

[xloem]

Thanks, that's very clarifying. Maybe I can try to stub an interface out with a naive implementation if I am bored, and submit a PR.

[romeric]

You are welcome to submit a PR if you wish

[romeric]

Also see #103