How does advanced spmv support $y = \alpha A x + \beta y$

[yhmtsai]

Does the current interface support advanced spmv like ($y = \alpha A x + \beta y$)?

AFAIK, it supported scaled like multiply(A, scaled(alpha, x), y) for $y = \alpha A x$ from https://github.com/SparseBLAS/spblas-example-project/blob/8ba01d8364171f3c888df31f4e95874c433a43db/src/simple_spmv.cpp#L18-L19
How does it apply to advanced spmv? something like multiply(A, scaled(alpha, x), scaled(beta, y)) or another function?
scaled(beta, y) reads the $beta * y$ not $y$. When putting the value back to the $scaled(beta, y)$, does y store value/beta?
If the scaled version also works on output, it also makes the original multiply(A, scaled(alpha, x), y) like $y = \alpha A x + y$
Another function for this advance spmv make more sense to me currently.

[BenBrock]

Sorry for being late getting back to you. I think the way we want to support this is with the four-parameter version of multiply.

// C = AB + E
multiply(A, B, E, C);

// three-parameter version is then essentially an alias to multiply(A, B, C, C)
// C = AB + C
multiply(A, B, C);

To set beta = 0, you'd then pass in scaled(0, C) for E.

// C = AB + 0*C
multiply(A, B, scaled(0, C), C);

Let me know if this is a priority for the examples and I can work on implementing the bits you need here. You can currently get a similar effect with

scale(0, C);
multiply(A, B, C);

[yhmtsai]

Yes, I currently use something similar to the last part to ensure C = AB
I feel that multiply(A, B, C) giving C=AB+C a bit weird from multiply meaning.

[BenBrock]

Fair point, let's discuss the desired behavior of multiply(A, B, C). (I'm now seeing that std::linalg's matrix_product does C = AB, so I think you're probably right.)