docs(perf): record real image deflate tradeoffs

This commit is contained in:
Rene Leonhardt committed 2026-10-19 13:14:37 +02:00
1 parent 4776d20b7a
commit a37198d4fa
1 file changed
+56 -40
+56 -40
View File
@@ -1,53 +1,69 @@
# Deflate backend comparison
Only the workspace flate2 backend changes from miniz_oxide to zlib-rs with runtime CPU detection. The 24 MP (6000×4000) RGBA8 cases are a repeating ramp and deterministic noise; inputs are prepared outside the timers.
Replace the workspace flate2 Rust backend with zlib-rs and runtime CPU detection. Tiles in .pcraft remain Zstandard; deflate changes its PNG previews.
Measured on 2026-10-19 against `d95fdf3181c01ad41e95456091a394b105a9afb2` on an Apple M4 Max
(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.
Retained with explicitly accepted tradeoffs. Some PSD ZIP files grow by up to 0.71%; the 16-bit city PSD import is about 6% slower in confirmation; the small web-layout PNG grows by 598 bytes (+1.40%); one .pcraft preview adds 25 bytes. Real photo PNGs become smaller and encode faster. Compression-effort probes recovered size only by introducing other speed regressions, so no effort change is applied. There is no native-format tile speedup from deflate.
| Workload | Main ms | This change ms | Time change |
|---|---:|---:|---:|
| ramp-png-encode | 42.114 | 43.989 | +4.5% |
| ramp-png-decode | 47.100 | 8.697 | -81.5% |
| ramp-psd-encode | 200.218 | 162.012 | -19.1% |
| ramp-psd-decode | 45.092 | 31.839 | -29.4% |
| ramp-pcraft-save | 82.126 | 83.149 | +1.2% |
| ramp-pcraft-load | 1.424 | 1.474 | +3.5% |
| noise-png-encode | 206.339 | 152.328 | -26.2% |
| noise-png-decode | 187.408 | 183.032 | -2.3% |
| noise-psd-encode | 1191.777 | 1002.312 | -15.9% |
| noise-psd-decode | 30.907 | 30.394 | -1.7% |
| noise-pcraft-save | 177.713 | 175.379 | -1.3% |
| noise-pcraft-load | 103.026 | 101.085 | -1.9% |
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.
| Encoded file | Main bytes | This change bytes |
|---|---:|---:|
| ramp-png-encode | 107,506 | 478,751 |
| ramp-psd-encode | 73,268,469 | 73,503,868 |
| ramp-pcraft-save | 88,188 | 88,214 |
| noise-png-encode | 82,731,323 | 82,272,418 |
| noise-psd-encode | 144,983,563 | 144,994,193 |
| noise-pcraft-save | 96,670,862 | 96,670,888 |
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.
Compressed bytes are not identical across backends. The ramp PNG gets larger; the noise case and decode timings must be considered alongside encode time. The native format stores Zstandard tiles and STORE ZIP entries, so this switch only changes its PNG previews; the repeating ramp deduplicates many tiles. No tile-compression speedup is attributed to deflate.
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.
The committed compatibility test compares old miniz_oxide and flate2 inflation of PNG IDAT and PSD ZIP/ZIP-prediction streams, checks exact decoded PNG/PSD pixels at supported depths, and compares complete .pcraft documents at U8/U16/F32 (tiles included) plus their PNG thumbnail. Float PSD composites are explicitly ZIP-encoded in the test because the IO exporter normally chooses Raw. The same test passed on the base and this branch.
| Document | Operation | Main ms | Branch ms | Time change | Paired 95% interval | Main / branch bytes |
|---|---|---:|---:|---:|---:|---:|
| canon-16 | psd-import (confirm) | 168.891 | 178.614 | +5.8% | +3.5..+5.9% | 140,673,004 / 140,673,004 |
| panasonic-8 | save | 166.659 | 162.322 | -2.6% | -3.9..-1.0% | 49,643,217 / 49,643,242 |
| panasonic-8 | load-common | 410.794 | 409.622 | -0.3% | -0.5..+0.3% | 49,643,217 / 49,643,217 |
| panasonic-8 | png-encode | 373.880 | 165.381 | -55.8% | -56.3..-47.4% | 24,569,965 / 24,238,416 |
| panasonic-8 | png-decode | 168.141 | 167.247 | -0.5% | -2.0..+1.0% | 24,569,965 / 24,569,965 |
| panasonic-8 | psd-export | 2938.276 | 1115.740 | -62.0% | -62.3..-61.6% | 98,823,859 / 98,837,585 |
| panasonic-8 | psd-import | 275.713 | 174.046 | -36.9% | -38.8..-34.7% | 98,823,859 / 98,823,859 |
| panasonic-16 | save | 357.998 | 355.900 | -0.6% | -1.8..+0.5% | 175,681,852 / 175,681,844 |
| panasonic-16 | load-common | 387.986 | 391.411 | +0.9% | -1.3..+1.1% | 175,681,852 / 175,681,852 |
| panasonic-16 | png-encode | 354.924 | 288.664 | -18.7% | -27.8..-14.2% | 111,583,441 / 109,241,690 |
| panasonic-16 | png-decode | 434.083 | 433.645 | -0.1% | -1.0..+1.7% | 111,583,441 / 111,583,441 |
| panasonic-16 | psd-export | 2645.838 | 2249.281 | -15.0% | -15.6..-13.8% | 274,409,683 / 275,654,405 |
| panasonic-16 | psd-import | 371.291 | 374.777 | +0.9% | -1.1..+1.8% | 274,409,683 / 274,409,683 |
| panasonic-32 | save | 639.677 | 631.556 | -1.3% | -5.2..+1.8% | 252,231,671 / 252,231,663 |
| panasonic-32 | load-common | 1338.160 | 1326.936 | -0.8% | -1.1..-0.2% | 252,231,671 / 252,231,671 |
| panasonic-32 | psd-export | 2522.700 | 1352.144 | -46.4% | -49.6..-44.6% | 461,042,978 / 459,545,262 |
| panasonic-32 | psd-import | 356.077 | 310.425 | -12.8% | -20.2..-4.7% | 461,042,978 / 461,042,978 |
| canon-8 | save | 89.139 | 85.965 | -3.6% | -8.9..+2.7% | 26,151,631 / 26,151,203 |
| canon-8 | load-common | 219.711 | 219.088 | -0.3% | -0.3..+1.0% | 26,151,631 / 26,151,631 |
| canon-8 | png-encode | 179.944 | 80.123 | -55.5% | -60.8..-40.0% | 13,720,268 / 13,551,023 |
| canon-8 | png-decode | 69.704 | 70.040 | +0.5% | -0.4..+0.6% | 13,720,268 / 13,720,268 |
| canon-8 | psd-export | 1074.094 | 475.802 | -55.7% | -56.1..-55.4% | 56,209,597 / 56,608,401 |
| canon-8 | psd-import | 142.523 | 96.323 | -32.4% | -34.4..-26.9% | 56,209,597 / 56,209,597 |
| canon-16 | save | 185.898 | 183.472 | -1.3% | -2.8..-0.2% | 88,750,003 / 88,749,555 |
| canon-16 | load-common | 193.409 | 193.704 | +0.2% | -0.2..+0.8% | 88,750,003 / 88,750,003 |
| canon-16 | png-encode | 169.432 | 135.364 | -20.1% | -26.2..-15.4% | 58,279,529 / 57,568,207 |
| canon-16 | png-decode | 211.947 | 211.829 | -0.1% | -0.4..+0.5% | 58,279,529 / 58,279,529 |
| canon-16 | psd-export | 1295.076 | 1130.697 | -12.7% | -13.5..-9.9% | 140,673,004 / 141,105,829 |
| canon-32 | save | 340.781 | 334.426 | -1.9% | -5.3..+0.7% | 120,957,984 / 120,957,536 |
| canon-32 | load-common | 774.663 | 778.401 | +0.5% | -1.3..+8.5% | 120,957,984 / 120,957,984 |
| canon-32 | psd-export | 1028.651 | 647.483 | -37.1% | -40.5..-36.1% | 228,234,238 / 227,497,106 |
| canon-32 | psd-import | 177.500 | 146.373 | -17.5% | -18.5..-17.1% | 228,234,238 / 228,234,238 |
| layout-masks | save | 18.588 | 18.268 | -1.7% | -2.7..+0.6% | 433,949 / 433,924 |
| layout-masks | load-common | 16.099 | 16.080 | -0.1% | -1.5..+0.6% | 433,949 / 433,949 |
| layout-masks | png-encode | 3.911 | 2.557 | -34.6% | -36.0..-32.3% | 42,842 / 43,440 |
| layout-masks | png-decode | 0.781 | 0.767 | -1.8% | -2.5..+2.2% | 42,842 / 42,842 |
| layout-masks | psd-export | 56.043 | 46.949 | -16.2% | -18.5..-15.9% | 493,006 / 493,145 |
| layout-masks | psd-import | 12.324 | 12.334 | +0.1% | -4.8..+2.3% | 493,006 / 493,006 |
| layout-mask-index | save | 7.367 | 7.202 | -2.2% | -5.5..-0.7% | 344,892 / 344,746 |
| layout-mask-index | load-common | 6.601 | 6.602 | +0.0% | -2.8..+4.7% | 344,892 / 344,892 |
| layout-mask-index | png-encode | 11.520 | 5.971 | -48.2% | -49.5..-46.0% | 115,229 / 114,892 |
| layout-mask-index | png-decode | 1.451 | 1.444 | -0.5% | -4.2..+0.5% | 115,229 / 115,229 |
| layout-mask-index | psd-export | 19.135 | 12.721 | -33.5% | -34.0..-31.5% | 411,435 / 411,594 |
| layout-mask-index | psd-import | 3.362 | 2.721 | -19.1% | -21.3..-11.2% | 411,435 / 411,435 |
Reproduce with the same release profile environment:
Every PNG encode decodes to identical source pixels, each PSD operation produces the same decoded layer/mask samples as main, and complete .pcraft documents compare equal at every depth. Compressed bytes may differ. The committed compatibility test independently inflates PNG IDAT and PSD ZIP/ZIP-prediction data using old miniz_oxide and checks .pcraft tiles plus thumbnails.
```sh
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 --test deflate_compatibility --no-run
# Run the printed executable at normal priority:
<printed-executable> benchmark_deflate_formats --ignored --nocapture --test-threads=1
```
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.
## Local validation
## 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-codecs -p photocraft-psd -p photocraft-format -p photocraft-io --locked`: passed (1956 passed, 9 ignored).
- `cargo clippy -p photocraft-codecs -p photocraft-psd -p photocraft-format -p photocraft-io --locked --all-targets -- -D warnings`: passed.