Vectorize per-channel PCA transform in run_for_all_spikes

[galenlynch]

Problem

_all_pc_extractor_chunk calls pca_model[chan_ind].transform() once per spike per channel in a Python loop:

for i in range(i0, i1):
    wf = traces[st - start - nbefore : st - start + nafter, :]
    for c, chan_ind in enumerate(chan_inds):
        w = wf[:, chan_ind]
        all_pcs[i, :, c] = pca_model[chan_ind].transform(w[None, :])

For a ~70-minute Neuropixels recording with ~10M spikes and ~26 sparse channels, this is ~260M individual sklearn transform calls, each on a 1×210 matrix. The Python/sklearn per-call overhead dominates.

Solution

Batch all spikes within a chunk by channel and call transform once per channel:

1. Extract all valid waveform snippets in the chunk at once using vectorized fancy indexing
2. Group spikes by channel index across all units
3. Call pca_model[chan_ind].transform(wfs_batch) once per channel with the full batch

In by_channel_local mode the PCA model is per-channel (not per-unit), so all spikes on a given channel share the same model regardless of unit identity.

Benchmarks

Synthetic (500 spikes, 10 channels, 50-sample waveforms)

	Time	Speedup
Original	0.126s	—
Vectorized	0.002s	53x

Results match: max absolute difference 9.5e-7 (float rounding).

Real data (Neuropixels probe 50213-3, 5 minutes of recording from local .dat)

• 706K total spikes
• 379 channels, 26 sparse channels per unit, 210-sample waveforms
• n_jobs=1, chunk_duration=10s

	Time	Per chunk	Speedup
Original	548.7s (9.1 min)	18.3s/chunk	—
Vectorized	110.3s (1.8 min)	2.1s/chunk	5.0x
Vectorized + numpy grouping	108.2s	2.0s/chunk	5.6x

Results match: max absolute difference 1.49e-08. np.allclose confirms identical output.

The real-data speedup is lower than synthetic because disk I/O, waveform extraction, and memory allocation are shared costs. The optimization only affects the transform call overhead, which is ~80% of chunk time in the original code.

Projected impact

For a full 69-minute recording, PC extraction drops from ~60 min to ~12 min.

RAM impact

Minimal. For a 10s chunk with ~25K spikes, the batch waveform array is ~25K × 210 × 4 bytes ≈ 20 MB per channel call, reused across channels. Peak additional memory vs. original: ~50 MB.

Changes

• _all_pc_extractor_chunk in principal_component.py: replaced per-spike per-channel loop with vectorized batch-by-channel approach
• No API changes, no new dependencies
• All existing test_principal_component.py tests pass
• A follow-up optimization replaces the per-spike Python dict loop (building the channel→spike mapping) with numpy unit grouping: loop over unique unit indices and use vectorized boolean masks. This reduces Python iterations from n_spikes × n_sparse_channels (~600K) to n_units × n_sparse_channels (~10K) per chunk.

Fixes #4485
Related: #979

[galenlynch]

The test failure is unrelated. Seems like a problem with cross-correlagram tests and fast_mode?