[Issue] When I set the convergence criterion to the absolute residual, the AMGX iteration log reports that “Relative residual has reached machine precision” and then stops further iterations.

[Haiqing-Fang]

When I set "convergence": "ABSOLUTE", "tolerance": 1e-7 in AMGX, the solver prints:

Relative residual has reached machine precision.

However, my stopping criterion is the absolute residual. Why is the relative residual being used? I would like the solver to use the absolute residual as the convergence criterion. Is it possible that my AMGX solver configuration is not actually taking effect?
I am using double precision, so the machine precision is about 1e-16.

AMGX solver configuration
CHECK_AMGX(AMGX_initialize());

AMGX_config_handle cfg;
CHECK_AMGX(AMGX_config_create_from_file(&cfg, "./amgx_SH.json"));

AMGX_resources_handle rsrc;
CHECK_AMGX(AMGX_resources_create_simple(&rsrc, cfg));

AMGX_matrix_handle A;
AMGX_vector_handle b, x;
AMGX_solver_handle solver;

CHECK_AMGX(AMGX_matrix_create(&A, rsrc, AMGX_mode_dDDI));
CHECK_AMGX(AMGX_vector_create(&b, rsrc, AMGX_mode_dDDI));
CHECK_AMGX(AMGX_vector_create(&x, rsrc, AMGX_mode_dDDI));
CHECK_AMGX(AMGX_solver_create(&solver, rsrc, AMGX_mode_dDDI, cfg));

CHECK_AMGX(AMGX_matrix_upload_all(
A, N, nnz,
1, 1,
row_ptr.data(),
col_ind.data(),
values.data(),
nullptr
));

CHECK_AMGX(AMGX_vector_upload(b, N, 1, rhs.data()));
std::vector h_x0(N, 0.0);
CHECK_AMGX(AMGX_vector_upload(x, N, 1, h_x0.data()));

CHECK_AMGX(AMGX_solver_setup(solver, A));
CHECK_AMGX(AMGX_solver_solve(solver, b, x));

AMGX_SOLVE_STATUS status;
CHECK_AMGX(AMGX_solver_get_status(solver, &status));
int iters = 0;
CHECK_AMGX(AMGX_solver_get_iterations_number(solver, &iters));

Configuration: amgx_SH.json:
{
"config_version": 2,
"determinism_flag": 1,

"solver": {
"solver": "FGMRES",
"scope": "main",

"preconditioner": {
  "solver": "AMG",
  "scope": "amg",

  "algorithm": "AGGREGATION",
  "cycle": "V",
  "max_levels": 30,
  "max_iters": 1,

  "smoother": "JACOBI_L1",
  "presweeps": 1,
  "postsweeps": 1,
  "coarsest_sweeps": 1,
  "selector": "SIZE_2"
},

"print_solve_stats": 1,
"obtain_timings": 1,
"monitor_residual": 1,

"max_iters": 1000,
"convergence": "ABSOLUTE",
"tolerance": 1e-7,
"norm": "L2"

}
}

Run log
Output around the end of the solve:
AMGX_solver_get_status(solver, &status):
669 16.2383 5.897965e-04 0.9862
670 16.2383 5.758887e-04 0.9764
671 16.2383 5.513113e-04 0.9573
672 16.2383 5.288395e-04 0.9592
673 16.2383 5.166605e-04 0.9770
674 16.2383 5.016301e-04 0.9709
675 16.2383 4.859083e-04 0.9687
676 16.2383 4.735401e-04 0.9745
677 16.2383 4.577987e-04 0.9668
678 16.2383 4.396150e-04 0.9603
679 16.2383 4.195545e-04 0.9544
Relative residual has reached machine precision
680 16.2383 4.045309e-04 0.9642
----------------------------------------------------------------------
Total Iterations: 681
Avg Convergence Rate: 0.9602
Final Residual: 4.045309e-04
Total Reduction in Residual: 9.873147e-13
Maximum Memory Usage: 16.238 GB
----------------------------------------------------------------------
Total Time: 13.1027
setup: 0.502298 s
solve: 12.6004 s
solve(per iteration): 0.0185028 s

[mattmartineau]

We have a check against the relative residual (but only for machine precision) inside the absolute convergence check. I understand your perspective, that the absolute residual only should be considered and that puts the contract on whoever chooses absolute residual to ensure that makes sense.

I'll fix this for the next release.