DPointNet: preserve projection rounding in fused current accumulation

[shixnya]

Summary

Fix the numerical boundary of DPointNet fused forward current accumulation: round the standalone synaptic projection to the compute dtype, then add the initial current exactly once.

• Fixed-four kernels no longer seed their reduction with the baseline; retain safe elementwise input-buffer forwarding.
• Atomic routes project into a separate zeroed output and add the initial current after projection on the same TensorFlow stream.
• Preserve the initial-current identity VJP, spike/canonical-weight VJPs, public ABI, accelerator selection and dynamics defaults.
• Add 92 GPU value/VJP/recomputation cases and native developer documentation.

This is independent of use_fused_recurrent_accumulation, the recurrent weight-gradient carrier; that option is not disabled by this fix.

Why

Seeding a projection with an existing current reassociates every synaptic contribution with the baseline, allowing small contributions to disappear and changing threshold crossings even in FP32.

The original full-V1 batch32 failure reproduced before backward or Adam: first current difference at tick 1, voltage difference at tick 2, and spike difference at tick 301. Independent legal FP32 reduction trees reproduce the early current values; emulation of the original fixed-four seeded/FMA arithmetic matches captured GPU outputs bitwise. This establishes a numerical-association mechanism, not an indexing/buffer-corruption diagnosis or exhaustive proof of race freedom.

The repair does not make atomic projection deterministic or promise bitwise equality for arbitrary regroupings of multiple projections.

Validation

The completed receipts below belong to isolated 8381b54c-based qualification. This PR applies the byte-identical CUDA source, regression tests and developer RST to current develop (b28e1cd, after #479). All native build inputs and the build script match the qualified snapshot. Integrated current-base GitHub CI results are separate and are not claimed green in advance.

Check	Result
New primitives, V10032	92 passed
Same 92 primitives, L40S	92 passed
Original fixed-four rounding regression	4 expected failures on original implementation, batches 1/4/16/32
Existing current/temporal-adjoint/precision-credit GPU subset, V10032	597 passed / 25 skipped
Complete CUDA-disabled TF2.21/Keras3 suite	1594 passed / 1074 skipped
Complete actual Python3.8/TF2.13/Keras2 suite	1594 passed / 1074 skipped
Full V1 FP32, batch32, 500 steps, checkpoint25, both families	Three updates pass exact counts and unchanged state/canonical-gradient/Adam tolerances, constraints and destructive all-master/input/optimizer/reset restore
Frozen batch1 oracle	All 32 initial per-neuron count vectors match exactly on both families
Canonical gradients	Worst repaired-vs-reference relative L2 below 7.15e-8; initial reference-vs-frozen-batch1 below 8.08e-8
Native docs	Full site rendered; 12 unchanged baseline diagnostics, none new
Publication checks	Python3.8/3.13 syntax, full patch review, whitespace and qualified-file hash checks passed

New tests cover FP16/FP32, batches 1/4/16/32, ordinary/fixed-four/active-row/device-queue/run-aggregation routes, eager/compiled repeated calls, distinct counted inputs, independent values, analytic VJPs and recomputation. The 92 GPU-only cases are included among CPU skips, not added to CPU pass totals.

Both native libraries were rebuilt for SM61/70/75/80/86/89/90 plus PTX90; execution here covers V10032 and L40S, not every compiled family. L40S full-V1 qualification retains native recurrent gradient accumulation ON; V100 auto selects it OFF.

Matched full-update timings (20 synchronized samples after 3 excluded warmups) are essentially unchanged against safe native current-accumulation-OFF controls: V10032 6.6161s vs 6.6384s (-0.34% time), L40S 3.0883s vs 3.0661s (+0.72%). No significant speedup is claimed.

Focused GPU reproduction

Rebuild both operators in an ABI-compatible environment, then run:

DPOINTNET_CUDA_ARCHS="61 70 75 80 86 89 90" \
  python -m bmtk.simulator.dpointnet.custom_ops.build
python -m pytest -q tests/simulator/dpointnet/test_current_accumulation_rounding.py

Limits

No complete GPU publication suite, repaired RTX3090 rerun, full-V1 mixed precision, distributed-training interaction, all-family or convergence qualification is claimed. Existing CPU CI disables fused CUDA and cannot replace hardware receipts. Local flake8 was unavailable; immutable environments were not modified.

Source-only fix: no generated binaries, private workload data, installed runtime migration or scientific-environment changes.

[Copilot]

🟡 Changes recommended

The regression suite does not detect omission of the intermediate FP16 projection rounding.

1 open issue

• Add FP16 cancellation test for intermediate projection rounding · New

What changed in this PR

Corrects DPointNet fused forward-current accumulation so projections round to the compute dtype before adding the initial current.

Changes:

• Separates projection reduction from baseline addition.
• Adds 92 GPU value, gradient and recomputation cases.
• Documents numerical guarantees and limitations.

File	Description
tests/​simulator/​dpointnet/​test_current_accumulation_rounding.py	Adds accumulation and gradient regressions.
docs/​autodocs/​source/​dpointnet_development.rst	Explains rounding boundaries and buffer handling.
bmtk/​simulator/​dpointnet/​custom_ops/​csr_spike_ops.cu.cc	Corrects fixed-four and atomic accumulation ordering.

_{Review effort: Balanced}