Make OpenCV performance comparisons equivalent

[bghira]

Summary

• compare resize against direct cv2.resize without unrelated RGB/BGR conversions
• compare luminance against direct COLOR_RGB2GRAY
• apply the same equivalent-work rule to the combined pipeline benchmark
• tighten resize output validation and report neutral relative timings
• clarify that TSR hides loops behind one Python call rather than eliminating them

Root cause

The competitive benchmarks charged two color conversions to OpenCV resize and an unnecessary RGB-to-BGR conversion to OpenCV luminance. The output then labeled the result a TSR advantage regardless of which implementation won. This obscured the actual resize bottleneck and overstated current performance.

Observed corrected result

On Apple M3 Max with the current main implementation and mixed image shapes:

• direct cv2 resize took about 0.31x TSR time
• OpenCV luminance took about 2.03x TSR time
• the direct OpenCV combined pipeline took about 0.76x TSR time

These are diagnostic point-in-time results, not assertions baked into the tests.

Validation

• corrected resize, luminance, and pipeline comparisons pass
• Python test suite passes
• repository formatting, lint, and competitive benchmark hooks pass

[Copilot]

Pull request overview

This PR adjusts the competitive performance benchmarks to ensure TSR vs OpenCV comparisons measure equivalent work (removing unrelated color conversions) and to report results neutrally without implying TSR always wins.

Changes:

• Update resize benchmarking to compare TSR against direct cv2.resize (no RGB↔BGR conversions) and tighten output validation.
• Update luminance benchmarking to compare TSR against cv2.cvtColor(..., COLOR_RGB2GRAY) directly (no extra RGB→BGR step).
• Apply the same “equivalent-work” rule and neutral relative-time reporting to the combined pipeline benchmark, plus clarify messaging about loops vs single Python calls.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.