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.
fix: thread a real per-channel weight stride through the field _jrls drivers - #53
Merged
Merged
Conversation
…drivers field_matvec_rls/diag_rls/relax_rls dispatch to the _jrls variants (matvec/diag/relax x absolute/membrane/bending, 9 functions) when the weight tensor's trailing dim equals the field's channel count (wc == nc, genuine per-channel weight). The driver functions here either didn't compute a per-channel stride at all, or (matvec_membrane_jrls) computed one but never passed it to the kernel call -- dead code, a leftover from an unfinished wiring attempt. Every channel silently got channel 0's weight instead of its own. Compute wsc = stride_wgt[nall] (the weight tensor's real last-dim stride -- the opposite of RLS/wc=1, which forces wsc to 0 to broadcast) and pass it through to the now-fixed _jrls kernels. part of fastfields-cpu-lib#65 _Workstream: claude-jitfields-to-fastfields_ Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016AjQcY78NgbagPSbPJRr6Z
part of fastfields-cpu-lib#65 _Workstream: claude-jitfields-to-fastfields_ Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016AjQcY78NgbagPSbPJRr6Z
Uh oh!
There was an error while loading. Please reload this page.
balbasty added a commit
that referenced
this pull request
Aug 3, 2026
…uperseded by the field_rls_is_jrls predicate fix (#59) * Revert "fix: thread a real per-channel weight stride through the field _jrls drivers (#53)" This reverts commit 24254d4. * Revert "fix: force the RLS weight-channel stride to 0 for broadcast reads (#48)" This reverts commit 1bfacc5. * chore: bump kernels submodule pin to the reverted _jrls per-channel stride part of fastfields-cpu-lib#65 * style: run clang-format on reg_field.h after the RLS/JRLS revert Mechanical formatting only (alignment, template-bracket spacing, line wrapping) -- no semantic change. The pre-#48 code this revert restores predates the clang-format lint gate and wasn't clean against it. --------- 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.
Summary
field_matvec_rls/diag_rls/relax_rlsdispatch to the_jrlsvariants(
matvec/diag/relax×absolute/membrane/bending, 9 functionshere) when the weight tensor's trailing dim equals the field's channel
count (
wc == nc, a genuine per-channel weight). The driver functionshere either didn't compute a per-channel stride at all, or
(
matvec_membrane_jrls) computed one but never passed it to the kernelcall — dead code, a leftover from an unfinished wiring attempt noted in
fastfields-cpu-lib#65. Every channel silently got channel 0's weight
instead of its own.
Fix
wsc = stride_wgt[nall](the weight tensor's real last-dimstride — the opposite of RLS/
wc=1, which forceswscto 0 tobroadcast) in all 9
_jrlsdriver functions and pass it through to thekernels.
kernelssubmodule pin to fix: give the field regulariser _jrls kernels a real per-channel weight stride fastfields-kernels#70,which gives the
_jrlskernels thewscparameter they were missing —without that fix this driver-side change alone doesn't compile.
Verification
Validated transitively through
fastfields-cpu-lib:make -C fastfields-cpu-lib test CXX=clang++— 12/12 suites, 44988 checks, 0failures — including the new ground-truth JRLS regression tests in
fastfields-cpu-lib's PR for this issue (which bumps this branch as itsimplpin).part of fastfields/fastfields-cpu-lib#65
Workstream: claude-jitfields-to-fastfields
🤖 Generated with Claude Code
Generated by Claude Code