Uh oh!
There was an error while loading. Please reload this page.
feat: bind field_relax (Gauss-Seidel relaxation for multi-channel fields) - #21
Merged
Conversation
…lds) `ff::field_relax` has existed in the C++ stack (cpu-lib dispatch, cuda-lib mirror, hub) since the field regulariser was ported, but it was never reachable from Python: `fastfields.dlpack` bound `flow_relax` only. That left the field family without the in-place solver its flow counterpart has, so the numpy/torch/cupy wrappers had no primitive to build `field_relax` on. Bind it exactly as `flow_relax` is bound, with the field family's per-channel penalty vectors (`absolute`/`membrane`/`bending` as length-C sequences) instead of the flow family's scalar Lame parameters -- `shears`/`div` are flow-specific and do not apply to independent field channels. Same in-place contract: `sol` is the warm start, mutated by `nb_iter` sweeps, and nothing is returned. The pinned `_fastfields_lib` already exports `ff::field_relax`, so no submodule bump is needed here. RLS/JRLS relax variants stay out of scope (a separate follow-up, per #33). part of fastfields/fastfields-lib#33 Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016AjQcY78NgbagPSbPJRr6Z
This was referenced Aug 1, 2026
Uh oh!
There was an error while loading. Please reload this page.
balbasty
commented
Aug 2, 2026
ContributorAuthor
Review summary (triage, jitfields-to-fastfields): merged as squash. 7/7 CI green. Binds Workstream: claude-jitfields-to-fastfields Generated by Claude Code |
balbasty added a commit
that referenced
this pull request
Aug 10, 2026
`field_relax` was the single binding in ext.cpp still declaring `int stream`; the other 47 (including its own sibling `field_relax_rls` and `flow_relax`) already take `intptr_t`. It was added in #21, in parallel with the widening pass in #22, and never converted. nanobind's `int` caster range-checks rather than truncating, so any genuine 64-bit CUDA stream handle -- what both Python callers pass, `torch.cuda.current_stream(...).cuda_stream` and `cupy.cuda.get_current_stream().ptr` -- was rejected with TypeError. `fastfields.torch.field_relax` / `fastfields.cupy.field_relax` therefore raised on essentially every real GPU call; latent only because there is no GPU in CI. Adds a regression test that calls every field_*/flow_* binding with a 1<<40 stream handle, plus a completeness guard so a new binding has to be added to the table. Refs fastfields/fastfields-lib#69 (item 1) Claude-Session: https://claude.ai/code/session_016AjQcY78NgbagPSbPJRr6Z 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 join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
What changed
Binds
ff::field_relaxasfastfields.dlpack.field_relax, plus a residual test for it.Why
ff::field_relaxalready exists all the way through the C++ stack —fastfields-cpu-libdispatches it (_field_relax+RX_DT),fastfields-cuda-libmirrors it, and the hub declares it inreg_field.h— but it was never reachable from Python: this layer boundflow_relaxonly. That left the field family without the in-place solver its flow counterpart has, so the numpy/torch/cupy wrappers had no primitive to buildfield_relaxon.The binding mirrors
flow_relaxexactly, with the field family's per-channel penalty vectors (absolute/membrane/bendingas length-Csequences) instead of the flow family's scalar Lamé parameters —shears/divdescribe cross-channel coupling of a displacement field and have no meaning for independent field channels. Same in-place contract:solis the warm start, mutated bynb_itersweeps, returningNone.The pinned
_fastfields_libalready exportsff::field_relax(itsreg_field.{h,cpp}are byte-identical to currentfastfields-libmain), so no submodule bump is needed here.RLS/JRLS relax variants are deliberately out of scope — a separate follow-up, per the "Remaining" list in #33.
How it was verified
Built against a clean
fastfields-libmainbuild (isolated tree,make libcpu && make all CXX=clang++→EXIT=0); symbol confirmed:pip install --no-build-isolation .succeeds and the op is callable:The new
test_field_relax_reduces_residualchecks against asym_matvec+field_matvec-derived reference: the residual(H + L)x - gmust shrink monotonically across six batches of 8 sweeps and end below1e-8, and the call must returnNone(in place).Independently cross-checked that the solver reaches the exact solution, not merely a small residual — building
(H + L)densely and comparing againstnumpy.linalg.solve:part of fastfields/fastfields-lib#33
Workstream: claude-jitfields-to-fastfields
🤖 Generated with Claude Code
Generated by Claude Code