Solving a linear system Ax = b where A, b are complex

[rtali-ia]

Hi There,

I am trying to solve a linear systems Ax = b where A,b are complex. I have saved A, b and x [the starting guess] as .mtx file. Here is my code

**using IndexType = int;
using vec = gko::matrix::Coo<std::complex>;
using mtx = gko::matrix::Coo<std::complex, IndexType>;

using cg = gko::solver::Cg<std::complex>;

auto A = gko::share(gko::read(std::ifstream("../data/A.mtx"), exec));
auto b = gko::read(std::ifstream("../data/b.mtx"), exec);
auto x = gko::read(std::ifstream("../data/x0.mtx"), exec);

auto solver_gen =
cg::build()
.with_criteria(gko::stop::Iteration::build().with_max_iters(20u))
.on(exec);

auto solver = solver_gen->generate(A);
solver->apply(b, x);**

I get the following error:

terminate called after throwing an instance of 'gko::DimensionMismatch'
what(): /usr/local/include/ginkgo/core/base/lin_op.hpp:304: validate_application_parameters: attempting to combine operators this [62476 x 62476] and x [1 x 62476]: expected matching row length
Aborted

I am not sure where is the dimension mismatch coming from? How do I fix the code?

Thanks

[upsj]

Your vector x0.mtx is stored as a row vector, it should be a column vector. You can either fix your output, or transpose it after reading:

x = gko::as<vec>(x->transpose());

[rtali-ia]

Thank you. But I am still receiving the error

error: ‘class gko::matrix::Coo<std::complex>’ has no member named ‘transpose’

Thanks for any inputs

[upsj]

Your vectors need to be of type matrix::Dense, not matrix::Coo

[rtali-ia]

Is there an example that I could use. I think I am struggling to deal with Complex vectors. Essentially, I pass a vector of size 62476 X 1 as a .mtx file to Ginkgo. This vector was created in python and converted to a .mtx file using mmwrite function. The problem is that when Ginkgo reads this .mtx file, it somehow converts itself into 1 X 62476 vector. I tried the instruction

x = gko::as<vec>(x->transpose());

but it still doesn't correct it. That is, it still prints the error

/usr/local/include/ginkgo/core/base/lin_op.hpp:303: validate_application_parameters: attempting to combine operators this [62476 x 62476] and b [1 x 62476]: expected matching inner dimensions

[rtali-ia]

Thank you Marcel. Please find the full code here. I have put the full folder structure here including the CMakeLists.txt . The main code is in simple-solver.cpp All the data (A.mtx, b.mtx, x0.mtx) has been placed in the data folder. You just need to follow the usual build protocol to reproduce the error. Thank you and much appreciate your eagerness to help me. Thank you again.

[MarcelKoch]

I don't have a box.com account. Please provide the code through some other means.

[rtali-ia]

No problem. Just created a public git repo. Thank you again. The repo is here

[MarcelKoch]

Thank you. I was able to run your code without any issues after commenting out the transposing of both x and b.

[rtali-ia]

Thank you Marcel. Really appreciate you taking the time to look into this. I am getting excellent performance from Ginkgo after being able to run it. I am not even using the CUDA kernels. It seems adding the CUDA kernels will speed up the performance to several orders of magnitude. Thank you again for your patience.

[rtali-ia]

Also Marcel, I see that you have an experimental direct solver too. Is there a sample code that you could share with me that uses your direct solver?