Feature: Smoothed Aggregation AMG with MIS-k selectors (prototype available)

[markadams]

Feature: Smoothed Aggregation AMG with MIS-k selectors (prototype available)

Overview

I have implemented Smoothed Aggregation (SA) AMG on top of AMGx's existing
aggregation framework. The goal is to bring AMGx closer to feature parity with
PETSc's GAMG solver for GPU-accelerated algebraic multigrid. A working prototype
is available at: https://github.com/markadams/AMGX/tree/feature/smoothed-aggregation-amg

What Was Added

1. SA Prolongator Smoothing

The standard SA step that transforms a tentative prolongator P_tent into a
smoothed prolongator P:

P = (I - ω D⁻¹ A) P_tent,   ω = (4/3) / ρ(D⁻¹A)

• Spectral radius ρ(D⁻¹A) estimated via power iteration (100 iterations)
• Supports scalar (block_size=1) and block problems
• Uses AMGx's existing SpGEMM infrastructure for the S * P_tent multiply
• Near-null space vector forwarded through the solver to the AMG hierarchy

Key files: src/aggregation/aggregation_amg_level.cu,
include/aggregation/aggregation_amg_level.h

2. Batched QR for Tentative Prolongator

Batched Modified Gram-Schmidt QR to construct P_tent from near-null space
vectors, one QR per aggregate.

Key file: src/aggregation/batched_qr.cu

3. MIS-k Aggregation Selectors

Two MIS-2 algorithms selectable via "mis2_algorithm":

Algorithm 0 (default): MIS-2 via 2× Galerkin coarsening loop — composes
two MIS-1 passes through a Galerkin coarse graph. Produces ~9.5 avg aggregate
size on 5-point Poisson.

Algorithm 1 (recommended): True MIS-2 directly on the original graph
without forming A²:

• Phase 1: each node checks all distance-2 neighbors; root if highest weight
• Phase 2: distance-1 assignment to nearest root
• Phase 3: propagation for remaining unassigned nodes

Produces ~7.1 avg aggregate size, matching PETSc GAMG's square-graph
aggregation quality.

Key file: src/aggregation/selectors/mis_selector.cu

4. Chebyshev Eigenvalue Reuse (Mode 4)

New chebyshev_lambda_estimate_mode=4 reuses ρ(D⁻¹A) computed during SA
prolongator smoothing, avoiding a redundant eigenvalue computation:

{
  "solver": "CHEBYSHEV",
  "chebyshev_polynomial_order": 1,
  "chebyshev_lambda_estimate_mode": 4,
  "chebyshev_lmin_denom": 11.0
}

lmin_denom=11 matches PETSc GAMG's default Chebyshev interval:
emax = 1.1*ρ, emin = 0.1*ρ = emax/11.

5. Bug Fix: Chebyshev solve_init zeros x when xIsZero=true

solve_init() now explicitly zeros x when xIsZero=true. Without this,
stale data from the previous outer PCG iteration accumulates via
x += gamma*p, causing divergence. (Also submitted as a standalone PR.)

Verification Results

100×100 2D Poisson (exact match with PETSc GAMG)

Using PETSc-exported aggregates imported into AMGx:

Metric	PETSc GAMG	AMGx SA
Iterations	71	71
P nnz	23,705	23,705
‖P_tent‖_F	3.780211634287129e+01	3.780211634287129e+01

AMGx SA matches PETSc GAMG exactly when given the same aggregates.

200×200 2D Poisson (40,000 DOFs)

	PETSc GAMG	AMGx SA (MULTI_PAIRWISE)
Levels	5	4
Iterations	14	23
Conv. rate	0.47	0.59
Grid complexity	1.17	1.15
Operator complexity	1.42	1.38

400×400 2D Poisson (160,000 DOFs)

	PETSc GAMG	AMGx SA (MIS-2, agg_levels=1)
Levels	5	5
Iterations	16	82
Conv. rate	0.49	0.86
L0→L1 ratio	7.1×	9.9×
Op complexity	1.43	1.24

The convergence gap at 400×400 is primarily due to aggregate size non-uniformity:
AMGx's parallel MIS with random hash weights produces aggregates up to size 26,
while PETSc's greedy sequential MIS caps at 13. Capping max aggregate size or
using a greedy ordering is the main remaining work item.

Configuration Example

{
  "config_version": 2,
  "solver": {
    "solver": "PCG",
    "preconditioner": {
      "solver": "AMG",
      "algorithm": "AGGREGATION",
      "selector": "MIS",
      "mis_k": 2,
      "aggressive_levels": 1,
      "smoother": {
        "solver": "CHEBYSHEV",
        "chebyshev_polynomial_order": 1,
        "chebyshev_lambda_estimate_mode": 4,
        "chebyshev_lmin_denom": 11.0,
        "preconditioner": {
          "solver": "BLOCK_JACOBI",
          "relaxation_factor": 1.0
        }
      },
      "coarse_solver": "DENSE_LU_SOLVER",
      "min_coarse_rows": 32
    }
  }
}

Current / Future Work

cuDSS Coarse Solver (in progress — not yet tested)

A GPU-native sparse direct coarse solver (CUDSS_SOLVER) is being developed to replace DENSE_LU_SOLVER. It uses NVIDIA's cuDSS library (CUDA 12.4+) for sparse LU or Cholesky factorization entirely on the GPU, avoiding the CPU round-trip. Code scaffolding is in place (src/solvers/cudss_solver.cu, include/solvers/cudss_solver.h) but has not yet been tested or validated.

Multi-GPU (MPI) Support (in progress — not yet tested)

Initial code scaffolding for multi-GPU is in place:

• The MIS-k selector includes halo exchange calls for distributed aggregation
• The cuDSS coarse solver includes a multi-GPU gather/scatter path (exact_coarse_solve=1)

None of this has been tested or validated in multi-GPU configurations. Full MPI testing on Perlmutter is a primary remaining work item.

Other Remaining Work

• Cap max aggregate size (target: max ≤ 13, match PETSc GAMG quality)
• Test block_size > 1 (elasticity problems with rigid body modes)
• Adaptive SA (multiple near-null vectors via randomized eigenvectors)

The full implementation with build instructions and verification scripts is at
https://github.com/markadams/AMGX/tree/main

[markadams]

Oh damn, I got confused with multiple remotes (gitLab vs gitHub) and I just pushed 62 commits from last week that were on my old remote at gitLab. This page was largely correct but updated also.