Classes
class	_DiagProxy

class	Case

class	Proxy

Functions
	settings_of (name)

	umean (s)

	emit (a, rec)

	common_rec (a, case)

	cmd_run (flow, a)

	g1_bed (a, case, conv, Ks)

	cmd_oracle (flow, a)

	cmd_g3 (flow, a)

	cmd_g5 (flow, a)

	cmd_g8 (flow, a)

	cmd_mpi (flow, a)

	main ()

Variables
	HERE = os.path.dirname(os.path.abspath(__file__))

tuple	BED_K_INF = (98.9156994, 98.9156995)

dict	G3_DEPTH = {"sphere": 1e-9, "bed": None}

Detailed Description

Gate instrument for the steady-march Anderson acceleration (doc/steady_acceleration.md, rev 2,
work orders WO-4, WO-5, WO-7 and WO-8). An INSTRUMENT, not a ctest: it prints one summary line per run and,
with --json, appends one JSON record per run. The numbers go to doc/steady_acceleration_log.md.

Subcommands (all on the module on PYTHONPATH; bound the OpenMP pool):

  oracle   WO-4: the C++ accelerator against the NumPy oracle (tests/study/anderson_oracle.py,
           rev 2) — the per-step relative velocity residual of unconditional step(True) calls on
           two identical solvers, max relative difference over the first --steps (30) steps.
  run      G1 / G2 / G6 / G7: march_to_steady per window (0 = accelerate=False), recording steps,
           accelerated steps, converged / reason, K, wall time, pressure iterations per step by
           phase (mixed vs plain), accelerator seconds, memory; with window 0 among the windows,
           the G7c line per case: plain converged vs accelerated converged per window. The tight
           staggered bed also prints the G1 bed line: |K_acc/K_inf - 1| <= |K_plain/K_inf - 1| +
           1e-8, K_plain from window 0 of the same run or --k-plain (the logged certificate).
  g3       G3: --steps (400) unconditional acc.step(True) per dt (--betas), tight; status,
           restarts, running-min residual, final residual and K. With --extend-to N (WO-8): past
           --steps, keep stepping until the running-min residual reaches the depth bar (sphere
           1e-9) or N steps in total; every criterion is evaluated at the end of the run. The bed
           has no depth bar: run it with --steps 3000 (a fixed 3000 calls per dt). K is sampled
           after the last call that did not restart (review R3).
  g5       G5: interrupt an accelerated march at step --at (25), checkpoint get_field u,v,w,p (or
           set_state, velocity only), restore into a fresh solver, march again; against the
           uninterrupted accelerated march.
  g8       G8: accelerator overhead (acc.seconds over --steps accelerated steps) against the mean
           plain step (last_step_timers()['step'], after --warmup plain steps, so both are timed
           late in the march), and memory_bytes against the §6.3 formula.
  mpi      G4: tight march_to_steady under mpirun (flow.mpi_block + init_mpi), global <u_x>
           monitor; and --hash: the state hash after --steps accelerated steps (np = 1 against a
           serial build).

Cases (as the oracle builds them):
  sphere  §11 of doc/collocated_invisible_subspace.md: one sphere at the centre of the periodic unit
          cube, phi (0.125), rho = F = 1, mu (1), nu dt / h^2 = beta (6) unless --dt; Stokes unless
          --advection. K = F L^3 / (6 pi mu R <u_x>) (Zick & Homsy normalisation). The Z&H simple
          cubic array is the same case at --phi 0.343 / 0.45.
  bed     a random arrangement (.npz: centres, radii, L), A1's K = F L^3 / (3 pi mu D <u_x> N_p).
          --force sets the body force (finite Re), --implicit-advection A1's implicit advection.

Settings: production = PCG(200, 1e-8), 200 velocity sweeps, rtol 1e-4, max_steps 5000; tight =
PCG(400, 1e-12), velocity residual tolerance 1e-12, rtol 1e-10, max_steps 20000.