How to deal with multiple right hand sides.

[g873249824]

Hi Professor, I have a basic question.
If I have vectors rhss and xs as arguments to solver operator () as (rhss,xs), which are multiple right hand sides (continusly stored) and solutions, I want to handle rhss and xs inside the header file by changing the operator ().
Can I declare a vector reference to part of rhss or xs without additional memory storage? or something alternative？

{
vector &rhs=rhss(0,n-1);
vector &x=xs(0,n-1);
//do some iteration cycles for the first rhs.
}
{
vector &rhs=rhss(n,2n-1);
vector &x=xs(n,2n-1);
//do some iteration cycles for the second rhs.
}// maybe i will continue  the iteration cycle for first rhs

How can i implement this in a elegant way for several backends?
With best regards.

[ddemidov]

Hey,

You should be able to use amgcl::make_iterator_range as in

for(int i = 0; i < nv; ++i) {
    auto b = make_iterator_range(&rhss[i*n], &rhss[i*(n+1)]);
    auto x = make_iterator_range(&xs[i*n], &xs[i*(n+1)]);
    solve(b, x);
}

[g873249824]

Thanks for your help.
I have just tried that, i think it works.
Sincerely！

[g873249824]

Sorry to bother you again, I was trying to do something like operator = or - like

vector t = b;
// or vector t= t-b;

I have tried converting it to numa_vector<rhs_type> and functions like backend::copy or backend::axby and I think I tried the wrong usage (because it is an iterator type). What is the right way to do this if I use 'amgcl::make_iterator_range'?

[ddemidov]

The following works for me:

#include <amgcl/backend/builtin.hpp>

int main() {
    std::vector<double> x(10, 1), y(10, 2);

    auto X = amgcl::make_iterator_range(x.begin(), x.end());
    auto Y = amgcl::make_iterator_range(y.begin(), y.end());

    // y = x
    amgcl::backend::copy(X, Y);

    std::cout << "1. y = " << y[0] << std::endl;

    // y = y - x
    amgcl::backend::axpby(-1, X, 1, Y);

    std::cout << "2. y = " << y[0] << std::endl;
}

1. y = 1
2. y = 0

[g873249824]

I will try that. Thanks for your time and patience.

[g873249824]

I don't know if I can implement this for several backends, such as Eigen backend or cuda.

[ddemidov]

• It should work for cpu-based backends (you can get pointers to the raw data from Eigen vector)
• I don't think cuda thrust library has range concept, so its not possible with the cuda backend (you would have to copy the current vector chunk into a smaller vector)
• It could probably work with vexcl backend using slices (https://vexcl.readthedocs.io/en/latest/expressions.html#slicing) but I have not tried this.

[g873249824]

Really thanks for your comment.
Maybe i will try to copy the data to a smaller vector so I can try to compile for different backends.