Add convergence-rate CI for WENO5/3/1 and MUSCL2#1404
Add convergence-rate CI for WENO5/3/1 and MUSCL2#1404sbryngelson wants to merge 42 commits intoMFlowCode:masterfrom
Conversation
With eps=5 the prim->cons covariance error O(eps^3 h^2) swamps WENO5's O(eps^2 h^5) error at all practical resolutions, giving apparent rate 2. Using eps=0.01 shifts the crossover to h~0.22 so that N=16..64 (h=0.63..0.16) are firmly in the WENO5-dominated regime and show rate ~5, as in Johnsen & Colonius (2006) who similarly used a weak vortex. - hcid=283: vortex strength now read from patch_icpp(patch_id)%epsilon (defaults to 5 if not set, preserving backward compatibility) - case.py: eps_vortex=0.01, c_max updated accordingly - run_convergence.py: resolutions 16,32,64; WENO5 tolerance 4.0 (rate>=4) - convergence.yml: updated to 16 32 64
CFL=0.02 eliminates RK3 temporal error so WENO5 shows rate 5. WENO3 expected rate is 2 (not 3) on smooth-extremum problems per Henrick et al. 2005 — JS smoothness indicators degrade at f'=0. The 2D vortex job remains the authority on WENO3 rate 3.
1D runner now uses examples/1D_euler_convergence/case.py (single-fluid,
density sine wave, no non-conservative alpha equation) instead of the
two-fluid advection case. The five-equation model non-conservative alpha
transport degrades unlimited MUSCL to ~1st order on two-fluid problems
even with identical EOS; the single-fluid Euler case gives clean rates.
Add muscl_lim=0 (unlimited central-difference) to MUSCL reconstruction:
- src/simulation/m_muscl.fpp: central slope = 0.5*(slopeL+slopeR)
- toolchain/mfc/params/definitions.py: add 0 to choices
- toolchain/mfc/case_validator.py: allow muscl_lim in {0..5}
1D and 2D convergence cases default to muscl_lim=0 for smooth problems
where TVD limiters would stall at smooth extrema. Runner default CFL=0.02
so RK3 temporal error O(h^3) stays negligible for WENO5 rate-5 verification.
N=16 (m=n=15) is below the minimum grid size for WENO5's 5-point stencil and causes pre_process to abort. Start all 2D tests at N=32. WENO3 on the stationary isentropic vortex is pre-asymptotic at N=32-128: fitted rate ~2.02, local rate 1.83->2.21 (converging toward 3). Lower threshold from 2.2 to 1.8 (expected=3, tol=1.2). The rate still clearly exceeds the 1D result (1.77, Jiang-Shu smooth-extremum degradation), which is the distinction this test is designed to demonstrate.
…d to 0.95 Default and CI resolutions: 128 256 512 1024. WENO5 is capped at N=512 per-scheme (hits double-precision floor at N=1024 after 111k steps; error ~2.6e-12 vs 4.2e-12 at N=512, rate collapses to 0.69). N=128-512 gives fitted rate 4.99. WENO1 threshold tightened from 0.6 to 0.95 (tol 0.05): pairwise rates 0.95->0.97->0.99 at N=128-1024, fitted 0.97. MUSCL2 shows exact rate 2.00 at all doublings N=128-1024.
WENO5 capped at N=512 (machine-precision floor kills rate at N=1024, error ~2.6e-12, rate collapses to 0.69). WENO3 starts at N=256 to skip coarse pre-asymptotic points. WENO1 and MUSCL2 use full [128,1024] range. Thresholds: WENO5>=4.8, WENO3>=1.8, WENO1>=0.95, MUSCL2>=1.9.
Tests are too slow (~30-45 min each job) to run on every PR push. Manual trigger via workflow_dispatch lets you run them deliberately when verifying numerics, not on every commit.
Claude Code ReviewHead SHA: 3a069d2 Files changed:
FindingsFortran-side checker for
|
Both 1D and 2D runners now test all 7 schemes: WENO5/3/1, MUSCL2, TENO5, WENO7, TENO7. Both case.py files gain --teno and --teno-ct flags. Resolution bounds: 1D: TENO5 [128,512], WENO7/TENO7 [128,256] (precision floor near N=512) 2D: TENO5 [32,128], WENO7/TENO7 [64,128] (MFC stencil constraint: N>=35) TENO CT values match the Shu-Osher examples: 1e-6 for order-5, 1e-9 for order-7. weno_eps tightened to 1e-40 (was 1e-16) for all WENO/TENO schemes in case.py.
Scripts import numpy which lives in build/venv, not the system Python.
Removed --no-build from run_case() in both runners. The generic ./mfc.sh build created a binary without the analytical IC (case.fpp), so --no-build would silently fall back to that binary and crash with SIGILL when the IC code wasn't found. MFC's hash system rebuilds only when case.fpp changes, so each scheme triggers one build then reuses the binary for all N values. Also drop WENO7/TENO7 tolerance to 4.0 (threshold >=3.0): local runs confirm rate ~3.7 due to RK3 temporal and spatial h^7 errors being comparable at N=128-256.
With CFL=0.02 (1D) or CFL=0.4 (2D) and RK3 time integration, the temporal error O(dt^3) is comparable to the O(h^7) spatial error at N=128-256, giving a spurious fitted rate of ~3.7 instead of 7. Fix: bake CFL=0.005 directly into WENO7/TENO7 extra_args. This drops the temporal error by (0.005/0.02)^3=1/64 in 1D and (0.005/0.4)^3=1/51200 in 2D, making spatial error dominate. All other schemes keep CFL=0.02 (1D) or 0.4 (2D). Threshold updated from >=3.0 to >=6.5 (tol 0.5) in 1D and >=6.0 (tol 1.0) in 2D. Also adds --cfl arg to 2D case.py so the runner can override it per scheme.
1D: use N=64,128 (not N=128,256). At N=256 the WENO7 spatial error (~1.7e-14) falls below the round-off accumulation floor (~2.5e-12 for ~28M cell-steps), giving a spurious rate of -0.21. At N=64 the spatial error is ~2.8e-10, well above the floor, and the rate between N=64 and N=128 is 6.97. Add N=64 to the CI resolution list accordingly. 2D: remove WENO7/TENO7 from the 2D test. The isentropic vortex has a primitive->conserved covariance floor O(eps^3*h^2). For WENO7 scheme error O(eps^2*h^7) to dominate requires h > eps^(1/5) = 0.40 (N < 25). At N=64-128 the covariance floor gives rate ~2, masking the 7th-order spatial rate regardless of CFL. The 1D test cleanly verifies WENO7/TENO7 7th-order convergence on a pure advection problem without this floor.
Threads num_ranks through run_case() and test_scheme() in both runners. GitHub ubuntu-latest runners have 4 vCPUs; using all 4 reduces simulation time for schemes with large Nt (e.g. WENO7/TENO7 at 27k-56k steps).
read_vf1_1d and read_cons_vf1 were reading only p_all/p0/, so with
num_ranks>1 only the first rank's N/num_ranks cells were included in
the L2 error. Now both readers iterate over p0..p{n-1} and concatenate.
For the L2 norm this is correct regardless of spatial ordering across
ranks. Validated: 4-rank WENO7 |vf0|=64 (not 16) and error matches
single-rank result exactly.
Tests WENO1/3/5/7, MUSCL-minmod/MC/VanLeer/SUPERBEE, TENO5/7 on the 1D Sod problem (shock + contact + rarefaction). Self-convergence: compare N vs 2N via cell-averaging, no exact Riemann solver needed. Rate ~1 for all schemes by Godunov theorem; WENO1 and SUPERBEE use wider tolerance (0.5) due to contact-dominated O(h^0.5) L1 diffusion at these resolutions. CI runs at 64/128/256/512 with 4 MPI ranks.
… from conservation check WENO5/TENO5 require (N/2) >= num_stcls_min*weno_order = 25 cells/rank in a 2x2 decomposition; N=32 only gives 16 cells/rank which fails MFC's decomposition check. Setting min_N=64 (32 cells/rank) fixes this. The isentropic vortex has zero net linear momentum by symmetry, making the relative conservation error formula ill-conditioned (divides near-zero by 1e-300). Density and energy have large nonzero integrals and are the meaningful conserved quantities to check.
MUSCL-SUPERBEE's pre-asymptotic L1 rate at N=64 is ~0.40 (contact-dominated), which pulls the fitted rate below the 0.5 threshold. Setting min_N=128 skips that pre-asymptotic point; the asymptotic rates (N=128,256,512) are ~0.56.
Previously triggered on every push to every branch, running expensive 1D/2D/temporal/Sod jobs (60-90 min total) on every dev push. Now triggers on: push to master, any pull_request, or workflow_dispatch.
…e, add traceback, validate output sizes, fix silent PASS on zero data
ⓘ You've reached your Qodo monthly free-tier limit. Reviews pause until next month — upgrade your plan to continue now, or link your paid account if you already have one. |
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## master #1404 +/- ##
==========================================
- Coverage 64.96% 64.94% -0.02%
==========================================
Files 72 72
Lines 18871 18872 +1
Branches 1571 1573 +2
==========================================
- Hits 12259 12257 -2
- Misses 5638 5639 +1
- Partials 974 976 +2 ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
|
Thanks @claude for the review. Responding to each finding: 1. Fortran-side 2. IC 283 Jensen's inequality covariance floor — the math is correct, but the convergence test measures 3. Workflow triggers without branch filter — already fixed; the current |
|
Claude Code is working… I'll analyze this and get back to you. |
Convergence tests now register as TestCases (kind='convergence') with a
'Convergence -> {1D,2D,Sod,Temporal} -> <scheme>' trace prefix, dispatched
through _handle_case to the runner library in toolchain/mfc/test/convergence.py.
Skipped by default; opt in via --only Convergence.
Drops the four standalone run_*.py runner scripts and the _convergence_common
helper module — their content moved into toolchain/mfc/test/convergence.py.
The convergence.yml workflow shrinks from 4 hand-rolled jobs invoking python
runners to a single ./mfc.sh test --only Convergence call.
The existing 2D vortex test can't measure 7th-order rate because the
primitive->conserved IC has an O(eps^3 h^2) covariance floor that dominates
WENO7's O(eps^2 h^7) scheme error at testable resolutions (h > eps^(1/5)
needed; eps=0.01 gives N < 25). Add a 2D smooth-advection case with rho =
1 + 0.2 sin(2pi(x+y)) and v=(1,1) on the unit square with periodic BCs;
linear primitives mean no covariance floor.
Cases registered: Convergence -> 2D Adv -> {WENO7, TENO7} at N=[80, 96],
CFL=0.005, 4 ranks, T=0.5 (one diagonal period). Both measure rate 6.86
(threshold 6.5) in ~5-8 min wall time per case.
Drop the 2D isentropic vortex case (eps=0.01 had a covariance floor that
forced wide tolerances 0.4-1.2 just to pass) and route all 2D convergence
schemes through the diagonal-advection case introduced for WENO7/TENO7.
Linear primitives → no covariance floor → tight 1D-like tolerances and
all 7 schemes (WENO1/3/5/7, MUSCL2, TENO5/7) on the same case.
Measured 2D rates (CFL=0.02 except WENO7/TENO7 at 0.005, T=0.5):
WENO5 4.99 (need >= 4.7) MUSCL2 2.01 (need >= 1.9)
WENO3 1.74 (need >= 1.7) TENO5 4.99 (need >= 4.7)
WENO1 0.84 (need >= 0.8) WENO7 6.86 (need >= 6.5)
TENO7 6.86 (need >= 6.5)
Suite wall time on 4 ranks: ~23 min (WENO7/TENO7 dominate at ~6/8 min).
Per-scheme resolution lists replace the min_N/max_N filter pattern (cleaner
when low-order schemes need different N than WENO7). Renames the runner
key '2d_vortex' → '2d_advection'.
Cell-shift mode in convergence runner: T = K*h per resolution, compare q(T) against np.roll(q(0), +K) in each dim (analytical solution for periodic linear advection with v=1). Cost is O(1) in N (Nt = K*c/CFL independent of resolution) instead of period mode's O(N) Nt. Forces num_ranks=1 for unambiguous Fortran-order data layout. Migrate 1D/2D advection schemes (WENO1/3/5, MUSCL2, TENO5) to cell-shift mode with K=1, expected_order = scheme_order + 1. WENO7/TENO7 stay in period mode since cell-shift signal hits machine precision at typical N (h^8 < 1e-15 at N=64). Add 3D advection case + 5 cell-shift cases (skip WENO7/TENO7). Empirical rates land cleanly across dimensionalities: WENO5 1D 6.00 2D 6.00 3D 6.00 (need >= 5.7) WENO3 1D 2.50 2D 2.49 3D 2.47 (need >= 2.2; effective 1.5 at extrema) WENO1 1D 2.00 2D 1.99 3D 1.97 (need >= 1.8) MUSCL2 1D 3.00 2D 3.01 3D 3.02 (need >= 2.7; unlimited central) TENO5 1D 6.00 2D 5.99 3D 5.99 (need >= 5.7) WENO7 1D 6.97 2D 6.86 (period mode; need >= 6.5) TENO7 1D 6.97 2D 6.86 (period mode; need >= 6.5) Wall: 1D 2.7 min, 2D 15.5 min (WENO7/TENO7 period dominates), 3D 14.5 min.
In cell-shift mode T = K*h scales with h, so accumulated truncation error T*h^p = h^(p+1) gives raw measured rate p+1 — one greater than the scheme's spatial order. The +1 is an artifact of scaling T with h, not an extra order of accuracy in the scheme. Subtract 1 from the displayed value so the runner reports the scheme's canonical spatial order p in both modes; the spec's expected_order field always means p directly. Output now self-documents: cell-shift: 'Fitted spatial order: 5.00 (need >= 4.7)' period: 'Fitted rate: 6.97 (need >= 6.5)'
Apply the toolchain conventions (BenchCase pattern, MFCException, typing.List) to convergence.py and trim newcomer-hostile bits: - Replace dict-based spec with @dataclasses.dataclass ConvergenceSpec — fields and types are visible at the top of the module instead of scattered across .get(default) calls inside each runner. - Drop the _RUNNERS string registry; spec.runner is now a Callable, so cases.py just stores the function reference directly (run_h_sweep, run_dt_sweep, run_sod_l1). - Drop io.StringIO + redirect_stdout; runners build a list of lines and '\n'.join it. No stdlib magic. - Rename runners to verb-noun (run_h_sweep / run_dt_sweep / run_sod_l1). - Use common.MFCException for run failures (matches the rest of toolchain). - Add a module docstring that names the three flavours up front. - Rename time_mode='cell_shift' string flag → cell_shift int (0=period). - Rename temporal-spec field 'N' → 'fixed_N' for clarity. - Collapse cases.py registration into one loop driven by a small table of (schemes, dim_label, case_path, ndim, cons_vars, cell_shift, ppn) tuples. Same numerical results pre/post (verified 1D suite + Sod WENO1 + Temporal RK1).
1 participant
Closes #305
Summary
Adds a convergence CI workflow (
.github/workflows/convergence.yml) with three jobs:1D spatial convergence (
run_convergence_1d.py) — single-fluid Euler density sine wave, one period advection; all schemes with 4 MPI ranks:2D spatial convergence (
run_convergence.py) — isentropic vortex eps=0.01 with Gauss-Legendre IC (hcid=283); N ∈ [32, 128], 4 MPI ranks:Temporal order (
run_temporal_order.py) — fixed N=512/WENO5, vary CFL with 4 MPI ranks:Also adds
examples/1D_euler_convergence/andexamples/2D_isentropicvortex_convergence/benchmark cases, and--time-stepperargument to the 1D case.Test plan