[Sad Catapella N6][Performance/Memory] RGB apply allocates multiple full float32 intermediates instead of a fused bounded-memory transform #53

Baseline / classification

Normalization Batch 6 at f784c95e3049483f87466150ffc5764375fa631d. Confirmed allocation pattern, peak RSS and speed improvement hypotheses require benchmarks. No tests executed and no source modified.

Evidence

• apply_rgb() lines 58–74 copies each input image to float32 and uses ((x-mi)/(ma-mi+eps)), creating at least subtraction/output temporary arrays in standard NumPy evaluation.
• Parameters.apply() lines 161–207 forwards all RGB patch pixels to apply_rgb; even when an input is already float32, astype(np.float32) defaults to copying, and a new result is required by tests/test_rgb.py lines 277–290.
• CSBDeep implementation optionally uses numexpr to evaluate the same fused float32 expression; this is a useful benchmark comparator, not a reason to add a dependency or presume parity/performance.

Optimization and Numba requirements

Benchmark a preallocated float32 output and sequential np.subtract(..., out=output), np.divide(..., out=output) to reduce whole-image intermediates while retaining the exact CSBDeep float32 operation order. Compare existing NumPy, preallocated ufuncs, optional numexpr, and only then a fused numba.njit kernel where dtype/layout trade-offs justify JIT startup. The return must never share memory with input (even float32), and RGB output must be bitwise identical to the direct CSBDeep reference—not just close. Include uint8/uint16, constant zero-scale, varied R/G/B values, noncontiguous C-first/C-last, RGB patches and full FOVs (e.g. 256² to 8192²), cold/warm compiled timing, peak RSS, allocations and threads. Do not silently substitute approximate arithmetic or new clipping semantics. Keep Numba optional only if measured improvements justify complexity; no benchmark result claimed.