Agent:claude-fastfields-to-teeny
Part of the tensor-native-boundaries umbrella (fastfields-lib#57), Phase B. One coordinated PR-set with the cpu-lib half (fastfields-cpu-lib "Phase B3" issue); this issue covers the impl repo (pushpull.h). Convention: TEENY-MIGRATION.md §9 — R2 spellings at.ndim / at.size(at.ndim - 1).
Tier-1 status confirmed (no kernel work here):kernels/pushpull/teeny.h's vox:: entry points are already tensor-native — vox::pull<D,O,B,reduce_t,offset_t>(VOut out, const VIn inp, const reduce_t loc[D], int extrapolate, bound_t rt) and siblings take teeny cell views; the loc[D] coordinate stays a plain array by design (§9.1: a vector of values, not a tensor operand — R2 does not apply to it). This issue is the impl-entry re-skin only.
Current signatures (exact; 4 entry points)
template <int D, int O, bound_t B, typenamereduce_t, typenamescalar_t, typenameoffset_t>
voidpull(offset_t nbatch, int extrapolate, bound_t bound,
scalar_t * out, constscalar_t * inp, constscalar_t * grid,
constoffset_t * size_out, constoffset_t * size_inp, constoffset_t * size_grid,
constoffset_t * stride_out, constoffset_t * stride_inp, constoffset_t * stride_grid)
(push identical; count drops the inp triple; grad adds bool ABS to the template and wraps out at nbatch + D + 2.) Each body opens with three tny::as_anyrank(..., rank, tny::copy_meta) calls over arrays the caller just exploded — pushpull already passes per-tensor size/stride arrays (no shared-size hazard), so this module is the cleanest mechanical case: the carriers replace 6–9 raw array arguments per entry with 2–3 self-describing ones.
Target
template <int D, int O, bound_t B, typenamereduce_t, classAO, classAI, classAG>
voidpull(AO ao, constAI ai, constAG ag, int extrapolate, bound_t bound)
// offset_t = decltype(ao.size(0)); nbatch = ao.ndim - D - 1 (R2)// nsp (spatial voxels per batch cell) = product of ag.size(nbatch + d), d in [0, D)
nbatch and all six/nine size_*/stride_* arrays disappear (R2/R3); _grid_spatial reads the grid carrier instead of size_grid.extrapolate/bound stay: runtime kernel parameters, not geometry. D/O/B/ABS/reduce_t stay template parameters (R1 — dispatch direction unchanged).- Read-only operands (
ai, ag; push's ai) are const-element carriers (R4). push/count's scatter target ao stays non-const (atomic accumulate). - The
cur_b batch-cell caching (peel the (*spln, C) volume once per batch cell, not per voxel — a mixed-radix decode saved nsp−1 times per cell) and the flat voxel-parallel scatter via anyAtomicAdd are unchanged, line for line. Only the three-line as_anyrank prologue moves out. grad's out carrier arrives at rank nbatch + D + 2 (the extra (C, D) axis) — its nbatch derivation is ao.ndim - D - 2; say so in a comment, since it is the one entry whose two carriers disagree in rank.
Scope
pushpull.h only, all 4 entry points. NOT kernels/pushpull/teeny.h (tier-1, already native), NOT kernels/gather.h, NOT the legacy pushpull/1d.h interpreter (Phase E), NOT cuda-impl (its own Phase B issue). Old signatures deleted, not deprecated — cpu-lib converts in the same PR-set.
Gate (Phase A's, restated)
- Untouched kernel-layer instantiations (
vox::pull/push/count/grad, _make_axis, gather_sep, _pull_rec/_push_rec/_grad_rec) byte-identical at -O2 wherever they exist as standalone symbols in cpu-lib's pushpull.o, both compilers. - Re-skinned driver loops: no new instructions inside the per-voxel loop — in particular the
cur_b cache must still elide the per-voxel peel_front_at<-(D+1)>; a diff that shows the batch-cell peel back inside the loop is a failed gate, not jitter. - Oracle:
test_pushpull326 checks (clang make test baseline, 2026-08-04), identical counts clang++ AND g++, true make clean between; whole suite green, unchanged counts; asan/ubsan on touched paths. - g++ nuance (Phase A finding): untouched-neighbor whole-TU inlining jitter acceptable; the g++ bar is kernels + driver loops + oracles.
Judgment call to document in the PR
The out/grid (pull/grad) — or inp/grid (push) — spatial-extent equality that CHECK_SAME_SPATIAL pins lib-side is what the shared flat voxel index relies on. The impl currently trusts it; with carriers a one-line debug assert per entry (ao.ndim == ag.ndim + the D spatial extents) is nearly free — decide and document, don't silently widen scope.
Agent:
claude-fastfields-to-teenyPart of the tensor-native-boundaries umbrella (fastfields-lib#57), Phase B. One coordinated PR-set with the cpu-lib half (fastfields-cpu-lib "Phase B3" issue); this issue covers the impl repo (
pushpull.h). Convention:TEENY-MIGRATION.md§9 — R2 spellingsat.ndim/at.size(at.ndim - 1).Tier-1 status confirmed (no kernel work here):
kernels/pushpull/teeny.h'svox::entry points are already tensor-native —vox::pull<D,O,B,reduce_t,offset_t>(VOut out, const VIn inp, const reduce_t loc[D], int extrapolate, bound_t rt)and siblings take teeny cell views; theloc[D]coordinate stays a plain array by design (§9.1: a vector of values, not a tensor operand — R2 does not apply to it). This issue is the impl-entry re-skin only.Current signatures (exact; 4 entry points)
(
pushidentical;countdrops theinptriple;gradaddsbool ABSto the template and wrapsoutatnbatch + D + 2.) Each body opens with threetny::as_anyrank(..., rank, tny::copy_meta)calls over arrays the caller just exploded — pushpull already passes per-tensor size/stride arrays (no shared-sizehazard), so this module is the cleanest mechanical case: the carriers replace 6–9 raw array arguments per entry with 2–3 self-describing ones.Target
nbatchand all six/ninesize_*/stride_*arrays disappear (R2/R3);_grid_spatialreads the grid carrier instead ofsize_grid.extrapolate/boundstay: runtime kernel parameters, not geometry.D/O/B/ABS/reduce_tstay template parameters (R1 — dispatch direction unchanged).ai,ag;push'sai) areconst-element carriers (R4).push/count's scatter targetaostays non-const (atomic accumulate).cur_bbatch-cell caching (peel the(*spln, C)volume once per batch cell, not per voxel — a mixed-radix decode savednsp−1times per cell) and the flat voxel-parallel scatter viaanyAtomicAddare unchanged, line for line. Only the three-lineas_anyrankprologue moves out.grad's out carrier arrives at ranknbatch + D + 2(the extra(C, D)axis) — itsnbatchderivation isao.ndim - D - 2; say so in a comment, since it is the one entry whose two carriers disagree in rank.Scope
pushpull.honly, all 4 entry points. NOTkernels/pushpull/teeny.h(tier-1, already native), NOTkernels/gather.h, NOT the legacypushpull/1d.hinterpreter (Phase E), NOT cuda-impl (its own Phase B issue). Old signatures deleted, not deprecated — cpu-lib converts in the same PR-set.Gate (Phase A's, restated)
vox::pull/push/count/grad,_make_axis,gather_sep,_pull_rec/_push_rec/_grad_rec) byte-identical at-O2wherever they exist as standalone symbols in cpu-lib'spushpull.o, both compilers.cur_bcache must still elide the per-voxelpeel_front_at<-(D+1)>; a diff that shows the batch-cell peel back inside the loop is a failed gate, not jitter.test_pushpull326 checks (clangmake testbaseline, 2026-08-04), identical counts clang++ AND g++, truemake cleanbetween; whole suite green, unchanged counts; asan/ubsan on touched paths.Judgment call to document in the PR
The out/grid (pull/grad) — or inp/grid (push) — spatial-extent equality that
CHECK_SAME_SPATIALpins lib-side is what the shared flat voxel index relies on. The impl currently trusts it; with carriers a one-line debug assert per entry (ao.ndim == ag.ndim+ the D spatial extents) is nearly free — decide and document, don't silently widen scope.