marsaev

Tracked internally: AMGX-56

It is possible that another application will change it's memory usage during runtime, so saving memory usage at AMGX startup wouldn't be the best solution. However considering that AMGX uses...

Just making sure i'm not missing anything. Before this draft, AMGX overrided calls to `cudaMalloc()` and `cudaFree()` with the functions in this file. Also it provided interface for `thrust::global_thread_handle::cudaMalloc` and...

You can see example of this API usage here: https://github.com/NVIDIA/AMGX/blob/main/examples/amgx_mpi_capi_cla.c#L513 And how input data is set up: https://github.com/NVIDIA/AMGX/blob/main/examples/amgx_mpi_capi_cla.c#L395-L436

Tested configuration was GMRES + Aggregation Multigrid + Jacobi solver. This is still an experimental feature and actually didn't get much traction. We don't have specific date where we will...

> However, the complier always fails when the system solver is built. Right, sometimes some code changes are required to enable complex support for specific component (i.e. solver).

Added task to provide complex support matrix for algorithms to documentation

Hi there, Sorry for _that much_ delayed response, do you need any guidance on AMGX?

Tracking internally: AMGX-51

What is the full configuration of the solver you use? What version of code do you use? Have you tried latest revision for `main` branch? `SIZE_2` selector is used for...