Does it support complex number

[ztdepztdep]

I want to use STRUMPACK to solve Ax=b, where b is complex vector. it seems strumpack doesn't support it?

[pghysels]

You need to also pass A and x as complex.

[ztdepztdep]

You need to also pass A and x as complex.

the matrix A is real, only the x and rhs are complex

[pghysels]

That is not supported, you need to also treat A as complex.

[xiaoyeli]

If only b (and x) are complex, it's more efficient to solve 2 real linear systems:
A x_r = b_r (real part of b)
A x_i = b_i (imaginary part of b)
Then, the final solution is x = x_r + i * x_i

Going into complex arithmetic is more expensive.

[ztdepztdep]

If only b (and x) are complex, it's more efficient to solve 2 real linear systems: A x_r = b_r (real part of b) A x_i = b_i (imaginary part of b) Then, the final solution is x = x_r + i * x_i

Going into complex arithmetic is more expensive.

Thank you very much. If matrix A is symmetry but not SPD, do we have a special setting for this condition? I use the defaults seeting
" CSPOptions opts;
opts.rel_tol = 1e-8;
opts.type = SP_TYPE_HSS;
"

[ztdepztdep]

I didn't find the cuda example in the example directory. is it still in test ? | | ***@***.*** | | ***@***.*** |

…

---- Replied Message ---- | From | X. Sherry ***@***.***> | | Date | 10/01/2022 07:56 | | To | ***@***.***> | | Cc | ***@***.******@***.***> | | Subject | Re: [pghysels/STRUMPACK] Does it support complex number (Issue #77) | If only b (and x) are complex, it's more efficient to solve 2 real linear systems: A x_r = b_r (real part of b) A x_i = b_i (imaginary part of b) Then, the final solution is x = x_r + i * x_i Going into complex arithmetic is more expensive. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>

[pghysels]

The solver will off-load work to the GPU using CUDA. But the interface takes all input/output matrix and vectors in host/CPU memory. So there is no CUDA code in the examples.
You need to configure strumpack through CMake to use CUDA, and if you want MPI code and GPU support, you also need to enable SLATE, see here https://portal.nersc.gov/project/sparse/strumpack/master/installation.html .
To enable/disable GPU offloading in the solver, see here: https://portal.nersc.gov/project/sparse/strumpack/master/GPU_Support.html

[pghysels]

We do not have special support for symmetry.
Also note that GPU off-loading is not supported when HSS is enabled.