From a8b6dd9f5a0a3a8c194f498e6fabdee06ab59eb7 Mon Sep 17 00:00:00 2001 From: Claude Date: Sun, 9 Aug 2026 17:43:47 +0000 Subject: [PATCH] chore: bump cpu/cuda pins to the RLS/JRLS kernel fixes and test coverage cuda: b8e5e9c -> 6ae52bd (fastfields-cuda-lib#46) Completes the propagation of fastfields-cuda-impl#41, which moved cuda-impl's kernels pin to main after it had fallen nine commits behind. The CUDA path was missing two field/flow regulariser correctness fixes the CPU path already had: fastfields-kernels#64 make_kernel_bending_rls rescaled its whole coefficient table by 0.25, running the membrane penalty at half strength whenever the bending-order RLS/JRLS kernel is built. Upstream measured max |matvec_bending - matvec_bending_rls(w=1)| of 1.8e+00 (2D) / 1.7e+00 (3D) before, ~1e-14 after. fastfields-kernels#52 guards the weight-map neighbour reads in the RLS/JRLS kernels (out-of-bounds read). With this the CPU and CUDA paths finally sit on the same kernels commit (b09b284). cpu: bf57a02 -> 13af367 (fastfields-cpu-lib#83) Adds ground-truth coverage for the wc == 1 (JRLS broadcast) direction, which had none: the #65 predicate bug mis-routed both weight-map layouts, but only wc == nc was regression-tested. Also unswaps the RLS/JRLS labels in cpu-lib's public header and test comments. No source changes here; both sub-libraries' own CI (including cuda-lib's 51-minute nvcc build) is green on the pinned commits. Co-Authored-By: Claude Sonnet 5 Claude-Session: https://claude.ai/code/session_016AjQcY78NgbagPSbPJRr6Z --- cpu | 2 +- cuda | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/cpu b/cpu index bf57a02..13af367 160000 --- a/cpu +++ b/cpu @@ -1 +1 @@ -Subproject commit bf57a02434bb2e7aabd1d989c22e5a4b7d673627 +Subproject commit 13af367e439d1c23caf6c7fe1cff3aa6c4d63404 diff --git a/cuda b/cuda index b8e5e9c..6ae52bd 160000 --- a/cuda +++ b/cuda @@ -1 +1 @@ -Subproject commit b8e5e9c8629a3c0078c0caf321201e910e9a6cc3 +Subproject commit 6ae52bdb4e19f18156205a68f85a9122550e73c9