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.
refactor(reg_flow): port kernel_*/matvec_all/diag_all/relax_* onto teeny - #33
Merged
balbasty merged 4 commits intoAug 1, 2026
Conversation
These were still on the legacy index2offset batch loop even though matvec/diag for the same three energies were already teenified. Peel each (*batch,*spatial) cell down to its batch origin (rank-nbatch+ndim carrier, peel_front_at<-ndim>) and write the kernel at the existing center_offset — no per-voxel spatial decode needed since these write one stencil per batch cell, not per voxel. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015xcQBMEdA5eQtTsx2wB3AC
…nto teeny matvec_all/diag_all are stencil-shaped like matvec_bending/diag_bending (same peel + batch-cell caching + manual spatial decode), just with the lame+bending combined kernel and its extra shears/div parameters. kernel_all/kernel_lame are batch-only+center, mirroring the kernel_absolute/membrane/bending pattern from the previous commit (peel_front_at<-ndim> over batch cells, write at center_offset). matvec_lame/diag_lame are intentionally left untouched: they aren't in cpu-lib's reg_flow.cpp dispatch table (only relax_lame_/kernel_lame are wired), so they're out of scope for this port. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015xcQBMEdA5eQtTsx2wB3AC
First relax port in the project (reg_field's relax sweeps are still broken dead code, per CLAUDE.md). Keeps the Gauss-Seidel / patch- coloring algorithm byte-for-byte (same patch1/matvec_membrane<isub>/ diag_membrane<set>/PosDef::relax_ calls, same n < 2*niter sweep count), but replaces the index2offset_v2/index2offset per-voxel decode with a peel: sol/hes/grd are wrapped once as rank-(nbatch+ndim) anyrank carriers (no trailing channel needed -- only the batch-cell base pointer matters), peeled per batch cell with the same re-peel-on- batch-change caching used by the matvec/diag stencils, then the spatial offset is folded by hand from the cached cell pointer. Iteration order over the flat index range differs from the original Fortran-layout index2offset_v2 decode (batch cells now vary slowly, matching the stencil convention, instead of batch dim 0 varying fastest), but this doesn't affect the result: patch coloring makes same-color voxels within one sweep mutually independent, so every voxel's Gauss-Seidel update is a pure function of data written in prior sweeps, not of loop order. checks: 830, failures: 0 unchanged across clang++, g++, and g++ -fsanitize=address,undefined. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015xcQBMEdA5eQtTsx2wB3AC
…eeny Same peel-based port as relax_membrane_ (previous commit), applied to the other three Gauss-Seidel / patch-coloring relax sweeps now wired in cpu-lib's dispatch table: relax_bending_ (patch3, kernelsize_bending, matvec_bending<isub>/diag_bending<set>), relax_lame_ (patch2, kernelsize_lame, matvec_lame<isub>/diag_lame<set>), and relax_all_ (patch3, kernelsize_all, matvec_all<isub>/diag_all<set>, the combined bending+lame kernel with its extra shears/div parameters). The Gauss-Seidel math, sweep counts (pow<ndim>(2|3)*niter), and PosDef::relax_ call are untouched -- only the index2offset_v2 per-voxel decode is replaced by the same batch-cell peel + cache used by relax_membrane_. Verified against the CPU library's reg_flow test suite (residual and self-adjointness checks, not just compile-and-run) on clang++, g++, and g++ -fsanitize=address,undefined -O1: checks: 854, failures: 0, matching the current fastfields-kernels-pinned baseline exactly (the shared kernels checkout picked up an unrelated upstream diag_bending sign fix mid-session, which is why the true current baseline reads 854 rather than the 830 quoted before that fix landed -- confirmed by building the untouched, pre-this-branch reg_flow.h against the same kernels checkout and getting the identical 854/0). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015xcQBMEdA5eQtTsx2wB3AC
balbasty
commented
Jul 30, 2026
CollaboratorAuthor
Agent: Generated by Claude Code |
balbasty
merged commit Aug 1, 2026
ac21632
into
claude/fastfields-teeny-refactor-js42id
1 check failed
Uh oh!
There was an error while loading. Please reload this page.
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.
Summary
kernel_absolute,kernel_membrane,kernel_bending,kernel_all,kernel_lame): these five were still on the legacyindex2offsetbatch loop despitematvec/diagfor the same energies already being teenified. Peeled each(*batch,*spatial)cell to its batch origin (rank-nbatch+ndimcarrier,peel_front_at<-ndim>) and wrote the stencil at the existingcenter_offset.matvec_all,diag_all): mirrorsmatvec_bending/diag_bending's established peel + batch-cell-caching pattern, with the combined Lamé+bending kernel and its extrashears/divparams.relax_membrane_,relax_bending_,relax_lame_,relax_all_): the first relax port in the project. Kept the algorithm byte-for-byte (samepatch1/2/3,matvec_X<isub>/diag_X<set>,PosDef::relax_, sweep counts). Wrappedsol/hes/grdonce as rank-(nbatch+ndim)anyrankcarriers and replaced theindex2offset_v2/index2offsetper-voxel decode with the same peel-and-cache-per-batch-cell idiom used by the stencil matvec/diag. Iteration order differs from the original decode, but this is immaterial: patch coloring makes same-color voxels in one sweep mutually independent.matvec_lame/diag_lame(not in cpu-lib's publicreg_flow.cppdispatch prior to this — see fastfields-cpu-lib#39, which wires them in), and the*_jrlsvariants (not wired anywhere).Test plan
make -C ../fastfields-cpu-lib test CXX=clang++— 854 checks, 0 failuresmake -C ../fastfields-cpu-lib test CXX=g++— 854 checks, 0 failuresg++ -fsanitize=address,undefined -O1build — 854 checks, 0 failuresreg_flow.hreproduces the same 854/0 baseline against the current kernels checkout (the diag cross-term fix in fastfields-kernels#32 landed mid-task and bumped the baseline from 830→854; this branch's changes reproduce it exactly, no regression or self-adjointness drift)Generated by Claude Code