Umbrella tracking issue for bringing the flow regulariser to full parity with the jitfields API (_regularisers_flows.py::__all__) across the whole stack (kernels → cpu-impl/cuda-impl → cpu-lib/cuda-lib → lib → bind-py → numpy/torch/cupy → fastfields.any).
While porting we found and fixed genuine correctness bugs shared with upstream jitfields (reported as balbasty/jitfields#5).
Delivered
Remaining
Status detail lives in fastfields-lib/MIGRATION.md. Per-repo PRs merging the delivered work link back here.
Workstream: claude-jitfields-to-fastfields
Umbrella tracking issue for bringing the flow regulariser to full parity with the jitfields API (
_regularisers_flows.py::__all__) across the whole stack (kernels → cpu-impl/cuda-impl → cpu-lib/cuda-lib → lib → bind-py → numpy/torch/cupy →fastfields.any).While porting we found and fixed genuine correctness bugs shared with upstream jitfields (reported as balbasty/jitfields#5).
Delivered
shears/div) cross-channel block was not self-adjoint at reflecting boundaries.shears/divthrough the entire stack (C++ + Python).flow_relax(Gauss-Seidel solver) through the entire stack.flow_precond/flow_forward— pure Python compositions, no new C++. numpy (reg: add flow_precond / flow_forward (numpy) fastfields-numpy#12), torch (reg: add flow_precond / flow_forward (torch) fastfields-torch#11), cupy (reg: add flow_precond / flow_forward (cupy) fastfields-cupy#8),fastfields.any(reg: dispatch flow_precond / flow_forward (+ flow_relax) in fastfields.any fastfields#17).flow_kernel— materialise the operator's Toeplitz stencil, full stack. See flow_kernel: materialise the flow regulariser's Toeplitz stencil across the stack #31. Surfaced + fixed a latentkernel_allchannel-stride bug (regularisers/flow: fix channel strides typed as reduce_t in kernel_all fastfields-kernels#28)._add/_sub/in-place variants offlow_matvec/flow_diag— pure Python accumulate forms (inp ± op(...)). numpy (reg: add flow_matvec/flow_diag _add/_sub/in-place variants (numpy) fastfields-numpy#14), torch (reg: add flow_matvec/flow_diag _add/_sub/in-place variants (torch) fastfields-torch#13, fresh forms differentiable), cupy (reg: add flow_matvec/flow_diag _add/_sub/in-place variants (cupy) fastfields-cupy#10),fastfields.any(reg: dispatch flow_matvec/flow_diag _add/_sub/in-place in fastfields.any fastfields#18).Remaining
_add/_sub/in-place variants offlow_kernel(niche — accumulate a freshly-materialised stencil into an existing buffer)weight=) reweighted variants of matvec/diagmatvec_lame_jrls+ CUDA mirror) — same bug class as the Lamé fix, deferred until flow-RLS is wired with a weighted symmetry-test oracleStatus detail lives in
fastfields-lib/MIGRATION.md. Per-repo PRs merging the delivered work link back here.Workstream: claude-jitfields-to-fastfields