Denis Demidov

You could also try to use vexcl backend with cuda (`make solver_vexcl_cuda`).

@AXIHIXA , could you please check if #250 solves the compilation issues with CUDA 12?

I don't have much experience with using scotch library, but this indeed looks like there were some compilation issues with scotch. Scotch is found in amgcl using this file: https://github.com/ddemidov/amgcl/blob/master/cmake/FindScotch.cmake...

If amgcl have troubles finding scotch, you can help it by defining environment variables SCOTCHDIR and PTSCOTCHDIR. These paths should contain scotch.h and ptscotch.h (possibly under scotch/ptcotch subfolders)

Please have a look at this tutorial, it has the example of using CUDA backend with amgcl: https://amgcl.readthedocs.io/en/latest/tutorial/poisson3Db.html

Block solvers in amgcl are using statically sized blocks, so there is no way to use dynamic block sizes. You could try to split the matrix into the regular and...

The CUDA backend does not support mixed precision (it needs things like mixed precision matrix-vector product, and CuSparse did not have those when the backend was implemented, not sure about...

> constructing the preconditioner is done on CPU, but why do we need to specify backend type for amg precondtioner? The preconditioner is constructed on the CPU and then moved...

Should probably be [`runge_kutta4< state_type, my_float >`](https://github.com/headmyshoulder/odeint-v2/blob/master/include/boost/numeric/odeint/stepper/runge_kutta4.hpp#L104).

Looks like the compiler does not support C++11.