perf(kpm): evaluate 128-byte chunked memory layout for Pyramid downsample

[kalwalt]

Summary

Evaluate (and possibly port) the 128-byte chunked memory access pattern from the C++ BoxFilterDecimate — but only if benchmarks show cache misses dominate on real-world image sizes.

Context

Follow-up to #130 (M8 step 1). The C++ BoxFilterDecimate in WebARKitLib/lib/SRC/KPM/FreakMatcher/detectors/pyramid.cpp processes images in 128-byte column chunks with separate remainder handling:

int num_c = src_width/128;
int num_r = src_width%128;

for (int i = 0; i < src_height_minus_1; i += 2) {
    // ... 128-byte column loop
    for (int j = 0; j < num_c; j++, src_ptr += 128, dst_ptr += 64) {
        HorizontalBoxFilterDecimate128(tmp1, src_ptr);
        HorizontalBoxFilterDecimate128(tmp2, src_ptr + src_step);
        VerticalBoxFilter64(dst_ptr, tmp1, tmp2);
    }
    // ... remainder
}

This is a 2013-era L1 cache optimization. PR #130 intentionally dropped it because:

1. The output bits are identical with or without chunking — it's purely a memory access pattern.
2. Modern LLVM auto-vectorizes the simple row-major loop well.
3. Adding it back requires ~50 lines of split + remainder handling.

This issue tracks whether to add it back, gated on benchmark evidence.

Depends on

• perf(kpm): add criterion benchmark for Pyramid downsample #131 (need baseline ns/iter numbers per size)
• Likely also perf(kpm): SIMD path for Pyramid downsample (AVX2/SSE4.1/wasm32) #132 (SIMD will probably hide any cache cost the chunked layout was working around)

Scope

• After perf(kpm): add criterion benchmark for Pyramid downsample #131 and perf(kpm): SIMD path for Pyramid downsample (AVX2/SSE4.1/wasm32) #132 land, profile the SIMD downsample on cache-relevant sizes (1080p, 4K).
• If L1d misses are a measurable fraction of cycles (perf stat -e cache-misses,LLC-misses on Linux; equivalent on Windows), prototype the chunked layout and measure.
• Decision criteria: port only if it shows ≥10% speedup over non-chunked SIMD on at least one supported target.

Acceptance criteria

• Profile data attached to the PR/issue.
• Either:
  • A PR adding chunked path with benchmark proof, or
  • This issue is closed wontfix with a one-line note: "modern Rust/LLVM handles the access pattern well enough; chunking not worth the complexity."

References

• PR feat(kpm): port BoxFilterPyramid8u to Rust (M8 step 1) #130 — original port (scalar, no chunking)
• perf(kpm): add criterion benchmark for Pyramid downsample #131 — benchmark baseline
• perf(kpm): SIMD path for Pyramid downsample (AVX2/SSE4.1/wasm32) #132 — SIMD path
• C++ reference: WebARKitLib/lib/SRC/KPM/FreakMatcher/detectors/pyramid.cpp lines 369–409 (BoxFilterDecimate)