perf: Combat perf improvements

[ilaykav]

• Closes # - Perf Improvement
• Tests included or not required because: all tests pass, Added combat benchmark
• Release notes not necessary because: internal perf improvement

Devloped on venv, ran benchamrk locally, main changes:

• Replaced np.dot(x, np.ones(...)) with numpy broadcasting ([:, np.newaxis])
• Extract .values from DataFrames before computation loops
  Benchmarked locally on 50K cells × 500 genes, 2-3 speedup on pp.combat

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 78.59%. Comparing base (d15aede) to head (f994269).
⚠️ Report is 1 commits behind head on main.
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #4070      +/-   ##
==========================================
- Coverage   78.60%   78.59%   -0.02%     
==========================================
  Files         118      118              
  Lines       12753    12745       -8     
==========================================
- Hits        10025    10017       -8     
  Misses       2728     2728              

Flag	Coverage Δ
hatch-test.low-vers	77.89% <100.00%> (-0.02%)	⬇️
hatch-test.pre	78.47% <100.00%> (-0.02%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
src/scanpy/preprocessing/_combat.py	88.18% <100.00%> (-0.70%)	⬇️

[ilaykav]

~1.5x-2x times faster on 3K-20K cells, measured locally on my Intel i7-8665U python 3.14

[flying-sheep]

hi, thanks! there are a few superficial issues.

About the general idea, why do you think that to_numpy() influences anything? For pandas arrays that already wrap underlying numpy arrays, the method should have basically no performance impact.

[scverse-benchmark]

Benchmark changes

Change	Before [d15aede]	After [d94816f]	Ratio	Benchmark (Parameter)
-	495±30μs	389±3μs	0.79	preprocessing_counts.FastSuite.time_log1p('pbmc68k_reduced', 'counts-off-axis')
-	110±0.8ms	76.3±1ms	0.7	preprocessing_log.CombatSuite.time_combat

Comparison: https://github.com/scverse/scanpy/compare/d15aedeb0300d04bb327333e30bebfc0aa37d87c..d94816fc8f252ffa9a4f919aaacc80dbbf505eb4
Last changed: Fri, 24 Apr 2026 08:00:37 +0000

More details: https://github.com/scverse/scanpy/pull/4070/checks?check_run_id=72836821961

[ilaykav]

You're right, the broadcasting itself (and the memory save from removing the np.ones allocations, not the .to_numpy()) is what actually made the impact, dropping the .to_numpy() change

[flying-sheep]

thank you!