Uh oh!
There was an error while loading. Please reload this page.
This repository was archived by the owner on Aug 20, 2026. It is now read-only.
test: check the RLS field operator reduces to the unweighted one at w == 1 - #59
Merged
Merged
Conversation
… == 1 The reg_field RLS/JRLS tests only asserted self-adjointness (`<L(w)x, y> == <x, L(w)y>`). That oracle is blind to the *magnitude* of the penalty coefficients: a term computed at the wrong strength stays perfectly symmetric, so a scaling bug passes the suite unnoticed. That is exactly what happened in fastfields-kernels#40, where `make_kernel_bending_rls` rescaled its whole kernel table by 0.25 -- right for the second-order (bending) coefficients, wrong for the first-order entries, which also carry a membrane contribution wanting 0.5. The membrane penalty ran at half strength whenever the bending-order kernel was constructed, and every existing test still passed. Add `run_2d_matvec_rls_unit_weight`, which pins magnitude with an independent oracle: with a uniform weight map `w == 1` the weighted operator must reduce *exactly* to the plain unweighted `field_matvec` at the same order, since w == 1 is the identity case of the RLS generalisation. Restricted to the boundary conditions whose implicit extension maps a constant weight map to itself -- DCT2 (reflects) and DFT (wraps). Zero would extend the weight map by 0 and DST2 by -w, so neither is an identity case for the weighted operator and neither belongs in this check. Covers absolute, membrane and bending orders, RLS (wc=1) and JRLS (wc=C). The same oracle, evaluated directly against the kernel templates, gives max |matvec_bending - matvec_bending_rls(w=1)| = 1.837e+00 before the kernels fix and 4.441e-16 after it (2D, membrane-only, bending-order kernel), while the absolute-only and bending-only cases are bit-for-bit unchanged at 0.0 and 1.066e-14 -- the symptom reported in fastfields-kernels#40 and its resolution. Requires the kernels fix (fastfields-kernels#64); these rows fail without it. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016AjQcY78NgbagPSbPJRr6Z
balbasty added a commit
that referenced
this pull request
Aug 1, 2026
Picks up fastfields-cpu-impl's kernels-pin bump (fastfields-kernels#40's membrane-in-bending-RLS scale fix, and kernels#52's OOB wget guard), plus the C++ lint/format gate. impl: b6eae48 -> 655abb9 Unblocks #59 (the regression test for the membrane-RLS fix), which needs this pin so `field_matvec_rls` actually routes through the corrected kernel construction. Claude-Session: https://claude.ai/code/session_016AjQcY78NgbagPSbPJRr6Z Co-authored-by: Claude <noreply@anthropic.com>
Uh oh!
There was an error while loading. Please reload this page.
balbasty added a commit
that referenced
this pull request
Aug 2, 2026
Pulls in fastfields-cpu-impl#48, which forces the RLS weight-channel stride (wsc) to 0 in matvec/diag/relax_rls for absolute/membrane/ bending instead of deriving it from the wgt tensor's real (size-1) last-dim stride -- the latter let wgt[wsc*c] walk off the wc=1 buffer into the next spatial voxel's weight for field channels c>0, which is what run_2d_matvec_rls_unit_weight (added in #59, the regression test for kernels#40) was catching as 29 mismatched elements in test_reg_field. Verified: `make -C . test CXX=clang++` -- all 12 suites pass, 0 failures. test_reg_field: 8440 checks, 0 failures (was 29 failing). test_reg_flow: 11415 checks, 0 failures (flow's RLS path was never affected -- see fastfields-cpu-impl#48 for why). Closes#62 Co-authored-by: Claude <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for freeto subscribe to this conversation on GitHub.
Already have an account?
Sign in.
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Companion test for fastfields/fastfields-kernels#64 (fastfields-kernels#40).
Why the bug got through
The
reg_fieldRLS/JRLS tests only asserted self-adjointness(
<L(w)x, y> == <x, L(w)y>). That oracle is blind to the magnitude of thepenalty coefficients — a term computed at the wrong strength stays perfectly
symmetric, so a pure scaling bug passes the suite unnoticed.
That is exactly what happened in fastfields-kernels#40, where
make_kernel_bending_rlsrescaled its whole kernel table by0.25— correctfor the second-order (bending) coefficients, wrong for the first-order entries,
which also carry a membrane contribution wanting
0.5. The membrane penalty ranat half strength whenever the bending-order kernel was constructed, and every
existing test still passed.
What this adds
run_2d_matvec_rls_unit_weight, which pins magnitude with an independentoracle: with a uniform weight map
w ≡ 1, the weighted operator must reduceexactly to the plain unweighted
field_matvecat the same order, sincew ≡ 1is the identity case of the RLS generalisation.Restricted to the boundary conditions whose implicit extension maps a constant
weight map to itself — DCT2 (reflects) and DFT (wraps). Zero would extend the
weight map by
0and DST2 by-w, so neither is an identity case for theweighted operator and neither belongs in this check.
Covers the absolute, membrane and bending orders, for both RLS (
wc=1) andJRLS (
wc=C).Verification
The same oracle evaluated directly against the kernel templates gives
max |matvec_bending − matvec_bending_rls(w≡1)|:i.e. the symptom reported in fastfields-kernels#40 and its resolution. The
absolute-only and bending-only rows are bit-for-bit unchanged.
Depends on fastfields/fastfields-kernels#64 — the new rows fail without that
fix (which is the point). Merge that first, or this PR's CI will be red until
the kernels submodule pointer picks it up.
Workstream: claude-jitfields-to-fastfields
🤖 Generated with Claude Code
Generated by Claude Code