perf(io): remove the PSD export inlining hint

crates/io/src/pixels.rs

@@ -158,7 +158,6 @@ pub fn zero_sample(s: SampleType) -> Vec<u8> {
 }
 
 /// Quantizes a normalized value to big-endian bytes.
-#[inline]
 pub fn encode_be(v: f32, s: SampleType, out: &mut Vec<u8>) {
     match s {
         SampleType::U8 => out.push((v.clamp(0.0, 1.0) * 255.0 + 0.5) as u8),

docs/psd-interleave-performance.md

@@ -1,35 +1,43 @@
 # PSD pixel buffer allocation
 
-Only PSD pixel shuffling and sample-codec inlining change. Interleave writes into the final buffer rather than allocating an intermediate buffer per band. The kernel processes 24 million pixels with four full planes at 8-bit, 16-bit and float depth.
+Write PSD interleaved pixels directly into the final buffer. Direct RLE output is already on main. The unrelated export-side sample-encoder inlining hint is removed locally.
 
-Measured on 2026-10-19 against `d95fdf3181c01ad41e95456091a394b105a9afb2` on an Apple M4 Max
+Whole-photo imports improve by about 4–7%, and the authored designs by about 20–26%. An added export-side inlining hint accompanied a separately confirmed 2.3% float PSD export slowdown (2440.168 → 2497.423 ms). Removing that one line leaves export source equal to main; follow-up float export is +0.4% and 16-bit export +0.9%, with both paired intervals crossing zero. These results supersede the much larger previous interleave-only percentages.
-(16 CPUs, 128 GiB), macOS arm64, rustc 1.99.1. Each branch starts directly at that
-base; none of the other performance splits is applied. Release uses optimization
-level 3, LTO off, 16 codegen units, debug info off and incremental compilation off.
-Builds ran at nice 10 with APFS-cloned warm caches; measurements ran at normal
-priority with no concurrent builds. Timings are medians of five samples after one
-warm-up, except the compositor's ten Criterion samples. These synthetic workloads
-are reproducible checks of this change, not a representative image corpus.
 
-| Workload | Main ms | This change ms | Time change |
+Measured on 2026-10-19 against main `d95fdf3181c01ad41e95456091a394b105a9afb2`, independently of the other splits. Apple M4 Max, 16 CPUs, 128 GiB, macOS arm64, rustc 1.99.1; release O3, LTO off, 16 codegen units, debug off. Builds used nice 10 and warm APFS-cloned caches; measurements used default priority without concurrent builds or profiling. Five alternating paired samples follow a warmup; confirmation and fix runs use nine pairs. Time change compares the two medians; intervals resample paired ratios. Negative time change means faster. Small differences are not automatically established improvements or regressions.
-|---|---:|---:|---:|
-| U8 | 20.493 | 5.097 | -75.1% |
-| U16 | 31.827 | 6.312 | -80.2% |
-| F32 | 41.915 | 11.305 | -73.0% |
 
-These are interleave-kernel timings, not whole PSD import/export timings. Direct RLE output is already on main and is not part of this PR.
+The full-resolution CC0 Pixls photographs are P1000475.RW2 (24 MP foliage centre crop, no resampling) and IMG_4059.CR2 (12 MP city), each converted to U8/U16/F32 with the ProPhoto RGB profile. F32 derives from 16-bit development, not HDR capture. MIT psd-tools masks.psd supplies a 640 x 1394 web layout with 22 imported layers and three masks; mask-index.psd supplies a 1000 x 867 illustrated card with a mask. The sources and hashes are pinned by the existing corpus. These are the available authored layered files, not a large layered-photo corpus. No generated noise, repeating ramps, enlarged charts or duplicated photo layers are used.
 
-The committed equivalence test uses the literal previous main-branch interleave implementation. It compares raw bytes across U8/U16/F32, empty and partial bands, missing/truncated planes, fill pixels and inversion. Existing deinterleave reference and big-endian sample-codec round trips cover the companion paths.
+The shared harness times in-memory operations. Input loading, image conversion, thumbnail rendering, hashes and equivalence checks are outside timers. Every save uses a fresh PcraftWriter and a real rendered thumbnail. Common reads use main-encoded files; own-output reads are separate. PSD exports use ZIP merged images at U8/U16, Raw merged images at F32, and normal RLE layer channels at U8/U16 / ZIP prediction at F32. Results do not measure filesystem throughput or browser performance.
 
-Reproduce with the same release profile environment:
+| Document | Operation | Main ms | Branch ms | Time change | Paired 95% interval | Main / branch bytes |
+|---|---|---:|---:|---:|---:|---:|
+| panasonic-32 | psd-export (fixed) | 2471.456 | 2481.877 | +0.4% | -0.1..+1.7% | 461,042,978 / 461,042,978 |
+| panasonic-16 | psd-export (fixed) | 2649.246 | 2673.071 | +0.9% | -5.2..+1.6% | 274,409,683 / 274,409,683 |
+| panasonic-8 | psd-import | 275.828 | 262.237 | -4.9% | -5.9..-4.1% | 98,823,859 / 98,823,859 |
+| panasonic-8 | psd-export | 2937.854 | 2915.929 | -0.7% | -11.5..+5.6% | 98,823,859 / 98,823,859 |
+| panasonic-16 | psd-import | 365.901 | 349.827 | -4.4% | -4.4..-2.6% | 274,409,683 / 274,409,683 |
+| panasonic-32 | psd-import | 360.706 | 334.838 | -7.2% | -12.6..-4.4% | 461,042,978 / 461,042,978 |
+| canon-8 | psd-import | 141.331 | 135.878 | -3.9% | -5.1..-3.6% | 56,209,597 / 56,209,597 |
+| canon-8 | psd-export | 1052.421 | 1054.472 | +0.2% | -0.6..+1.3% | 56,209,597 / 56,209,597 |
+| canon-16 | psd-import | 177.327 | 169.952 | -4.2% | -5.5..-1.7% | 140,673,004 / 140,673,004 |
+| canon-16 | psd-export | 1248.449 | 1239.159 | -0.7% | -1.6..+0.6% | 140,673,004 / 140,673,004 |
+| canon-32 | psd-import | 175.721 | 163.271 | -7.1% | -9.3..-4.9% | 228,234,238 / 228,234,238 |
+| canon-32 | psd-export | 1008.772 | 1005.356 | -0.3% | -1.8..+0.7% | 228,234,238 / 228,234,238 |
+| layout-masks | psd-import | 12.509 | 9.981 | -20.2% | -20.8..-14.5% | 493,006 / 493,006 |
+| layout-masks | psd-export | 55.113 | 54.714 | -0.7% | -2.3..+1.2% | 493,006 / 493,006 |
+| layout-mask-index | psd-import | 3.299 | 2.443 | -25.9% | -30.8..-23.4% | 411,435 / 411,435 |
+| layout-mask-index | psd-export | 18.200 | 18.397 | +1.1% | -0.1..+1.7% | 411,435 / 411,435 |
 
-```sh
+Real PSD exports remain byte-identical and imported layer/mask float samples match main. The committed literal previous-interleave oracle covers depths, bands, fills, truncation and inversion.
-CARGO_PROFILE_RELEASE_LTO=false CARGO_PROFILE_RELEASE_CODEGEN_UNITS=16 CARGO_PROFILE_RELEASE_DEBUG=0 CARGO_INCREMENTAL=0 nice -n 10 cargo test --release -p photocraft-io --lib --no-run
-# Run the printed executable at normal priority:
-<printed-executable> benchmark_interleave --ignored --nocapture --test-threads=1
-```
 
-## Local validation
+Current local checks: affected `pixels::` unit tests (five passed, one ignored manual benchmark); strict all-target IO Clippy; affected IO wasm32 check. All passed. Follow-up exports at float and 16-bit depth remain byte-identical to main; no synthetic timing test was run.
+
+Local raw evidence and the shared harness: `.local/perf-real-261019/` in the repository checkout (ignored). This comparison supersedes the earlier synthetic performance table. Small synthetic unit fixtures remain correctness checks; ignored synthetic timing tests were not run.
+
+## Earlier split validation
+
+These checks passed when the split was prepared; they were not all rerun for this benchmark review. Current affected checks are listed above where production code changed.
 
 - `cargo test -p photocraft-io --locked`: passed (617 passed, 6 ignored).
 - `cargo clippy -p photocraft-io --locked --all-targets -- -D warnings`: passed.