Skip to content

benchmarks.md

Latest commit

 

History

History
839 lines (766 loc) · 66.6 KB

benchmarks.md

File metadata and controls

839 lines (766 loc) · 66.6 KB
Raw

Benchmarks

Every number here was measured on real camera files with the tools in this repository, and can be reproduced with them. FlashRaw is compared, on the same files and the same machines, with the open-source RAW decoders people use today, each built from source (or installed the way its users install it):

  • LibRaw, the most widely used RAW library (C++), decode and its dcraw_process() development to RGB;
  • the LibRaw + Little CMS pipeline: LibRaw develops to a linear wide-gamut image and the Little CMS colour-management engine converts it to sRGB, the usual way to get colour-managed output from LibRaw (Little CMS is not a RAW decoder: it is measured as that pipeline's colour stage);
  • RawSpeed, darktable's decoder (C++, raw data only);
  • rawler, the Rust decoder of dnglab, decode and its RawDevelop pipeline;
  • rawpy, the Python binding of LibRaw (what a Python user gets from pip install rawpy);
  • dcraw, the C program LibRaw grew out of.

Headline

Median speedup of FlashRaw over the 62 files of benches/popular-cameras.tsv (other library's time ÷ FlashRaw's on the same file; > 1 = FlashRaw faster), its geometric mean, and the number of files where FlashRaw is faster, out of the files both decode. dcraw has no threads: its one-thread time is used in the all-cores rows.

Comparison threads 🐧 median 🐧 geomean 🐧 FlashRaw faster 🍎 median 🍎 geomean 🍎 FlashRaw faster
decode vs LibRaw all 7.87× 6.20× 57/59 11.4× 9.11× 59/59
decode vs LibRaw 1 2.90× 2.40× 57/59 1.81× 2.02× 57/59
decode vs RawSpeed all 2.15× 2.00× 46/49 1.50× 2.00× 38/49
decode vs RawSpeed 1 1.29× 1.27× 41/49 0.83× 0.85× 13/49
decode vs rawler all 2.95× 2.61× 58/59 2.66× 3.08× 58/59
decode vs rawler 1 1.66× 1.71× 57/59 1.50× 1.54× 56/59
decode vs rawpy all 8.37× 6.74× 58/59 14.3× 13.6× 59/59
decode vs rawpy 1 3.08× 2.61× 58/59 1.85× 2.07× 57/59
decode vs dcraw all 14.5× 14.3× 42/42 29.1× 27.6× 42/42
decode vs dcraw 1 3.72× 3.99× 40/42 3.98× 4.24× 40/42
RAW→RGB vs LibRaw all 8.81× 7.91× 59/59 15.4× 14.1× 59/59
RAW→RGB vs LibRaw 1 2.06× 1.88× 59/59 1.77× 1.74× 59/59
RAW→sRGB vs LibRaw + Little CMS all 15.0× 15.0× 59/59 21.8× 19.9× 59/59
RAW→sRGB vs LibRaw + Little CMS 1 12.8× 12.2× 59/59 10.6× 11.0× 59/59
RAW→RGB vs rawler all 6.56× 6.42× 58/59 6.67× 5.81× 58/59
RAW→RGB vs rawler 1 4.70× 4.38× 58/59 3.61× 3.55× 58/59
RAW→RGB vs rawpy all 9.50× 8.72× 59/59 20.5× 17.3× 59/59
RAW→RGB vs rawpy 1 2.20× 2.09× 59/59 1.77× 1.75× 59/59
RAW→RGB vs dcraw all 9.98× 9.17× 42/42 23.0× 19.0× 42/42
RAW→RGB vs dcraw 1 1.98× 1.81× 40/42 1.95× 1.87× 40/42
RAW→RGB, FlashRaw GPU vs LibRaw all – – – 17.9× 17.1× 52/52

Bar chart: decode time relative to FlashRaw

Bar chart: RAW to sRGB time relative to FlashRaw

Dot plot: per-camera decode and RAW to RGB times

  • With all cores FlashRaw is the fastest decoder on almost every file. It decodes in parallel even the codecs whose data is one sequential stream (lossless JPEG, Nikon, Pentax, Hasselblad, Panasonic v8, Sony ARW1), which the others decode on one thread; LibRaw parallelises only its CR3, Fuji-compressed and Panasonic v8 decoders.
  • RAW → sRGB is where the gap is largest: FlashRaw develops in one fused, multithreaded pass, where LibRaw (and rawpy, which is LibRaw), dcraw and rawler run several full-image passes (LibRaw's and dcraw's mostly on one thread). Colour management with Little CMS adds yet another full-image pass to a LibRaw pipeline (see the stages).
  • On one thread the decoders are much closer, and RawSpeed is often as fast or faster, especially on Apple Silicon; see where FlashRaw is not the fastest.
  • Charts: charts/ (generated by tools/bench/charts.py, numbers in charts/data.json). Raw data and the full per-machine reports: benches/results/.

What is measured

Library Version, build decode RAW → RGB
FlashRaw this repository, version 0.1.0 (the development tree at the time of the runs; the commit the results were produced from was rewritten out of the history, so the exact state is not recorded — reproduce with the commands below), cargo build --release -p flashraw-cli --features jpegxl (+ gpu on macOS) flashraw benchmark --output raw-u16 --backend simd: open + full sensor decode to 16-bit samples, CPU --output rgba8 --backend simd: decode + bilinear demosaic + as-shot WB + colour matrix + sRGB 8-bit RGBA + orientation, one fused pass
FlashRaw GPU (macOS) same, --features gpu – --backend gpu --output rgba8: CPU decode (or GPU unpack for packed data), Metal develop, image downloaded to host memory
LibRaw 0.22.0, its Makefile.dist, -O3 (+ -march=native on x86-64), zlib; all cores = OpenMP build, 1 thread = default build open_buffer() + unpack() + dcraw_process() (user_qual=0 bilinear, use_camera_wb=1, output_bps=8, sRGB) + dcraw_make_mem_image()
LibRaw + Little CMS LibRaw as above + Little CMS 2.19.1 (static, -O3) – LibRaw unpack() + dcraw_process() to 16-bit linear ProPhoto RGB (bilinear, camera WB, gamm = 1, 1, no_auto_bright) + dcraw_make_mem_image(), then an lcms2 transform (linear ProPhoto → sRGB IEC 61966-2.1, perceptual, TYPE_RGB_16 → TYPE_RGBA_8); profiles and transform created per image. All cores: the transform runs over 32-row blocks with OpenMP (a shared transform is reentrant); 1 thread: one cmsDoTransform call
RawSpeed develop c835b05, CMake Release, WITH_OPENMP=ON, WITH_JPEG=OFF (+ -march=native on x86-64) RawParser::getDecoder() + checkSupport() + decodeRaw() + decodeMetaData() no RGB pipeline
rawler (dnglab) 0.8.0 from crates.io, cargo build --release, rayon rawler::decode() of an in-memory source: sensor data + metadata + RawDevelop::default().develop_intermediate() (scaling, PPG demosaic, WB, camera matrix, crop, sRGB gamma, f32) + conversion to 8-bit RGB
rawpy 0.27.1 wheel from PyPI (bundles LibRaw 0.22.1), CPython 3.13 (Linux), 3.9 (macOS) rawpy.imread(BytesIO) + raw_image (unpack) + postprocess(demosaic_algorithm=LINEAR, use_camera_wb=True, output_bps=8)
dcraw 9.28, cc -O3 -DNODEPS (+ -march=native on x86-64), one thread (it has no threading) dcraw -c -D -4 FILE > /dev/null (raw sensor values written as 16-bit PGM) dcraw -c -q 0 -w FILE > /dev/null (bilinear, camera WB, 8-bit sRGB PPM)
  • Every value is the median of 7 timed runs after one warm-up, in a fresh process per (file, library, mode, thread count). The input file is read into memory before timing (RawSpeed gets a padded copy, as its own reader provides) — except dcraw, a program rather than a library: each of its runs is a whole process (start-up, reading the file from the page cache, writing the image to /dev/null), so its times include a few ms the others do not pay.
  • Memory is the peak resident set size (ru_maxrss) of that process: the input file held in memory, the library's working buffers and its output. The process decodes one file 8 times in a row, freeing everything between decodes, so the peak is that of one decode. rawpy includes the Python interpreter and NumPy (about 27 MB before anything is decoded).
  • All cores = each library's default parallelism: FlashRaw's thread pool, OpenMP for LibRaw, RawSpeed and the lcms2 transform, rayon for rawler; rawpy's wheels use OpenMP on Linux but not on macOS. One thread: RAYON_NUM_THREADS=1, OMP_NUM_THREADS=1, LibRaw's default build. dcraw runs on one thread in both columns.
  • Files a library does not decode (error, or unsupported format) are left out of its comparisons. dcraw 9.28 "succeeds" on some files it cannot decode and writes an all-zero image (Canon CR3, whose decoder is an empty stub, and the Fujifilm compressed RAFs of current bodies): the script checks dcraw's output and counts those files as unsupported.

What the outputs are

The libraries do not produce byte-identical results, so the comparison is of equivalent work, not identical output:

  • decode: every library decodes the full sensor buffer. FlashRaw and LibRaw report it whole, masked (optical black) borders included; RawSpeed reports a view cropped to the camera's active area. All return black/white levels and white balance. FlashRaw applies DNG opcode lists by default (bad pixels, gain maps, …); RawSpeed applies a subset of list 1, the others none.
  • RAW → sRGB: FlashRaw, LibRaw, rawpy and dcraw demosaic bilinearly with the camera white balance and convert to 8-bit sRGB; rawler uses its PPG demosaic (more work than bilinear) in 32-bit floats. LibRaw, rawpy and dcraw also run an automatic-brightness histogram and write 3-channel RGB with a BT.709-style gamma; FlashRaw writes RGBA with the sRGB curve, applies DNG opcode list 3 and the orientation. The LibRaw + Little CMS pipeline skips the auto-brightness (it develops linear data) and writes RGBA, like FlashRaw.

Reproducing

python3 tools/corpus/fetch.py                        # CC0 corpus (raw.pixls.us)
cargo build --release -p flashraw-cli --features jpegxl     # add gpu for the GPU column

# FlashRaw alone (regression tracking):
python3 tools/bench/run.py benches/benchmark-set.tsv --iterations 5 \
    --out benches/results/<host>.jsonl --markdown benches/results/<host>.md

# Against the other libraries: see the header of tools/bench/compare.py for how to get and
# build them (nothing is vendored); then
tools/bench/compare.py build --libraw-src <LibRaw> --rawspeed-src <rawspeed> --rawspeed-build <build> \
    --lcms-src <Little-CMS> --dcraw-src <dcraw.c> --rawler-version 0.8.0 --rawpy rawpy==0.27.1
tools/bench/compare.py run benches/popular-cameras.tsv \
    --libs flashraw libraw rawspeed rawler rawpy dcraw lcms --out pass1.jsonl [--max-busy 0.15]
tools/bench/compare.py run benches/benchmark-set.tsv --libs flashraw libraw rawspeed --out set.jsonl
tools/bench/compare.py merge pass1.jsonl pass2.jsonl set.jsonl --out benches/results/<host>-compare.jsonl
tools/bench/compare.py report benches/results/<host>-compare.jsonl --popular benches/popular-cameras.tsv \
    --markdown benches/results/<host>-compare.md
tools/bench/charts.py                                # docs/charts/*.svg + data.json

cargo bench -p flashraw --bench kernels              # micro-benchmarks (criterion)
cargo bench -p flashraw-gpu --bench gpu_pipeline     # GPU (needs an adapter)

Machines and conditions

🐧 Linux x86-64 🍎 macOS Apple Silicon
CPU AMD Ryzen 7 9700X, 8 cores / 16 threads, 32 MB L3 Apple M1 Ultra, 16 performance + 4 efficiency cores
Memory, OS 64 GB DDR5, Linux 6.12 (Debian 13) 64 GB, macOS 27.0
Compilers rustc 1.99, GCC 14.2 rustc 1.95, Apple clang 17.0
Threads (all cores) 16 20
Conditions shared with other jobs (builds, browser tests; load average 2–15): before each measurement the script waited up to 20–60 s for other processes to use under 15 % of the CPU (they did for 97 % of the kept values, median 1.3 %); one full pass, a second pass over 37 of the 62 popular cameras, and every measurement taken while the machine was busier measured again; lowest median kept desktop session with background services (a virtual machine, storage indexing) using about 10–20 % of the CPU (top: 77–92 % idle); two full passes of the popular cameras, one of the 159-file set, lowest median kept; the two passes agree within 0.8 % (median) / 3.2 % (90th percentile)
Results linux-x86_64-compare.md macos-arm64-compare.md

Popular cameras (Apple M1 Ultra)

Times in ms, all cores, bold = fastest. ² dcraw has no threads (one-thread time). One-thread tables: result file.

Decode

Camera Format · compression MP FlashRaw LibRaw RawSpeed rawler rawpy dcraw ²
Canon EOS R5 (RAW) CR3 · canon-crx 46.7 80.9 138.0 ❌ 237.3 444.4 ❌
Canon EOS R5 (C-RAW) CR3 · canon-crx 46.7 100.7 141.0 ❌ 196.5 368.0 ❌
Canon EOS R5 Mark II (RAW) CR3 · canon-crx 47.9 83.2 141.4 ❌ 234.0 451.7 ❌
Canon EOS R6 Mark II (RAW) CR3 · canon-crx 25.5 45.7 76.9 ❌ 136.2 235.8 ❌
Canon EOS R6 Mark III (RAW) CR3 · canon-crx 34.0 59.6 100.4 ❌ 144.6 321.8 ❌
Canon EOS R7 (RAW) CR3 · canon-crx 33.7 60.0 100.6 ❌ 170.9 317.6 ❌
Canon EOS R8 (RAW) CR3 · canon-crx 25.5 45.4 75.3 ❌ 130.7 240.1 ❌
Canon EOS R50 (RAW) CR3 · canon-crx 25.1 44.8 75.3 ❌ 123.2 230.8 ❌
Canon EOS 90D (RAW) CR3 · canon-crx 33.7 31.2 99.7 ❌ 111.6 315.7 ❌
Canon EOS 5D Mark IV (RAW) CR2 · lossless-jpeg 31.3 20.5 431.4 110.2 251.5 428.3 587.6
Canon EOS 6D Mark II (RAW) CR2 · lossless-jpeg 27.0 17.6 389.7 121.2 238.2 379.3 541.9
Canon EOS Rebel T7 / 2000D (RAW) CR2 · lossless-jpeg 24.7 19.0 397.3 121.8 230.1 383.9 527.9
Nikon Z8 (14-bit lossless) NEF · nikon-huffman 45.7 44.0 544.4 218.3 356.5 570.7 857.1
Nikon Z9 (14-bit lossless) NEF · nikon-huffman 45.7 48.2 499.0 186.5 332.7 515.6 792.7
Nikon Z6III (lossless) NEF · nikon-huffman 24.5 25.0 284.6 110.5 186.2 297.6 452.9
Nikon Zf (lossless) NEF · nikon-huffman 24.5 24.9 267.5 95.9 174.1 279.5 426.7
Nikon Z7II (lossless) NEF · nikon-huffman 45.7 45.4 488.2 168.9 320.4 499.5 775.7
Nikon Z50 (lossless) NEF · nikon-huffman 20.9 22.0 246.9 96.3 157.7 261.8 389.9
Nikon D850 (14-bit lossless) NEF · nikon-huffman 45.7 45.1 487.7 170.4 324.9 504.4 769.8
Nikon D750 (14-bit lossless) NEF · nikon-huffman 24.3 24.8 255.8 84.8 166.3 263.2 413.2
Sony A7 IV (lossless) ARW · sony-lossless 36.7 19.2 413.6 11.7 38.7 427.2 ❌
Sony A7 IV (compressed) ARW · sony-arw2 33.0 5.86 87.4 8.77 8.38 85.0 260.1
Sony A7R V (lossless) ARW · sony-lossless 64.7 33.2 760.9 24.2 61.3 765.4 ❌
Sony A1 (compressed) ARW · sony-arw2 50.2 8.14 132.0 13.4 11.8 127.5 401.3
Sony A7 III (compressed) ARW · sony-arw2 24.3 4.49 64.7 7.20 6.43 62.5 198.7
Sony A7C II (compressed) ARW · sony-arw2 33.0 5.83 87.1 8.55 8.28 85.9 267.8
Sony A9 III (compressed) ARW · sony-arw2 24.3 4.65 63.6 6.83 6.82 62.9 197.3
Sony A6700 (lossless) ARW · sony-lossless 30.7 16.9 373.3 13.0 42.2 374.6 ❌
Sony A6400 (compressed) ARW · sony-arw2 24.3 4.47 63.0 6.54 6.41 62.2 202.5
Sony ZV-E10 (compressed) ARW · sony-arw2 24.3 4.47 64.7 6.60 6.40 64.0 203.0
Fujifilm X-T5 (lossless compressed) RAF · fuji-compressed 40.9 35.5 150.6 44.4 93.9 1167.2 ❌
Fujifilm X-H2 (compressed) RAF · fuji-compressed 40.9 46.5 172.8 ❌ 127.0 1477.7 ❌
Fujifilm X-S20 (lossless compressed) RAF · fuji-compressed 26.5 28.3 127.0 36.3 81.0 816.4 ❌
Fujifilm X100VI (lossless compressed) RAF · fuji-compressed 40.9 35.6 154.5 44.5 94.7 1178.0 ❌
Fujifilm X-T4 (uncompressed) RAF · uncompressed 26.7 1.60 18.8 1.18 3.26 16.7 160.5
Fujifilm GFX100 II (lossless compressed) RAF · fuji-compressed 103.4 67.3 267.2 76.6 143.1 2614.4 ❌
Panasonic Lumix S5II RW2 · panasonic-v8 24.1 15.1 131.1 62.4 119.1 238.6 234.5
Panasonic Lumix GH6 RW2 · panasonic-v8 25.2 17.0 133.2 63.5 117.0 251.2 253.0
Panasonic Lumix G9II RW2 · panasonic-v8 25.1 15.3 137.9 58.7 122.9 273.5 249.7
Panasonic Lumix S5 RW2 · panasonic-v7 24.3 2.61 8.48 0.60 3.79 8.58 241.4
Panasonic Lumix S1R RW2 · panasonic-v6 47.6 8.46 122.8 8.63 10.7 120.6 464.6
Panasonic Lumix GH5 RW2 · panasonic-v4 20.6 4.68 91.0 4.77 7.02 91.2 204.0
OM System OM-1 Mark II ORF · olympus 20.4 115.8 381.5 162.2 225.0 385.4 149.0
OM System OM-5 ORF · olympus 20.5 126.0 350.9 152.2 220.9 351.0 147.2
Olympus E-M1 Mark III ORF · olympus 20.5 122.2 339.7 147.6 197.9 339.1 443.1
Pentax K-3 Mark III PEF · pentax-huffman 26.6 29.8 204.3 83.6 173.5 292.7 448.3
Pentax K-1 Mark II PEF · pentax-huffman 36.6 38.0 293.4 131.2 257.3 406.4 639.9
Ricoh GR III DNG · lossless-jpeg 24.8 18.1 316.2 96.2 214.5 331.7 477.1
Leica Q2 DNG · uncompressed 47.4 4.04 292.9 49.2 4.54 366.0 674.9
Leica SL2 DNG · uncompressed 47.4 4.09 289.9 48.6 4.70 371.6 681.0
Leica M10-R DNG · lossless-jpeg 41.0 28.9 559.0 197.3 380.5 553.8 845.6
Hasselblad X2D 100C (compressed) FFF · hasselblad 105.3 76.6 1012.8 451.1 ❌ 1108.5 2216.0
Phase One IQ4 150MP (IIQ L) IIQ · phaseone 154.0 42.1 1006.4 ❌ 341.0 1058.7 4147.3
Sigma fp L DNG · lossless-jpeg 60.5 28.0 761.2 22.4 40.9 763.0 1183.4
Apple iPhone 12 Pro (ProRAW) DNG · lossless-jpeg 12.2 15.4 469.0 ❌ 27.8 465.2 577.1
Google Pixel 8 Pro DNG · lossless-jpeg 50.1 23.1 655.7 18.9 34.2 645.2 996.0
Google Pixel 6 Pro DNG · lossless-jpeg 12.5 6.45 163.5 4.92 10.6 163.3 254.0
Samsung Galaxy S23 Ultra DNG · lossless-jpeg 49.9 128.7 1916.4 549.6 1134.9 1803.2 2277.8
Samsung Galaxy S22 (JPEG XL) DNG · jpeg-xl 12.0 272.5 ❌ ❌ 172.6 ❌ ❌
DJI Phantom 4 Pro DNG · uncompressed 20.0 1.22 17.0 0.99 1.36 17.0 142.4
GoPro HERO11 Black (GPR) DNG · gopro-vc5 27.1 39.8 ❌ 18.5 ❌ ❌ ❌
GoPro HERO10 Black (GPR) DNG · gopro-vc5 23.3 35.9 ❌ 18.0 ❌ ❌ ❌

RAW → sRGB

Camera FlashRaw LibRaw LibRaw + Little CMS rawler rawpy dcraw ² FlashRaw GPU
Canon EOS R5 (RAW) 109.3 736.8 1104.5 491.8 1112.9 ❌ 99.1
Canon EOS R5 (C-RAW) 129.4 738.9 1116.2 438.1 1028.6 ❌ 119.7
Canon EOS R5 Mark II (RAW) 111.7 740.0 1122.9 510.2 1135.1 ❌ 103.0
Canon EOS R6 Mark II (RAW) 61.0 406.5 617.0 265.1 596.6 ❌ 56.8
Canon EOS R6 Mark III (RAW) 80.4 538.4 815.7 349.9 804.0 ❌ 73.9
Canon EOS R7 (RAW) 80.3 538.8 818.9 363.0 803.4 ❌ 74.2
Canon EOS R8 (RAW) 60.8 414.6 612.5 251.9 606.1 ❌ 55.8
Canon EOS R50 (RAW) 60.3 405.2 610.7 266.1 588.5 ❌ 56.3
Canon EOS 90D (RAW) 52.7 543.5 822.0 294.4 810.2 ❌ 45.5
Canon EOS 5D Mark IV (RAW) 39.8 837.5 1100.4 414.8 881.3 965.0 31.2
Canon EOS 6D Mark II (RAW) 33.7 744.3 968.6 384.6 764.2 857.4 28.4
Canon EOS Rebel T7 / 2000D (RAW) 34.3 719.3 931.9 369.3 750.9 824.5 28.3
Nikon Z8 (14-bit lossless) 72.4 1155.6 1541.7 601.2 1262.6 1296.1 63.7
Nikon Z9 (14-bit lossless) 72.4 1118.3 1484.4 581.4 1206.7 1251.3 63.0
Nikon Z6III (lossless) 41.5 628.9 824.4 320.0 668.4 681.6 36.5
Nikon Zf (lossless) 41.9 609.3 819.7 314.4 642.9 668.4 35.8
Nikon Z7II (lossless) 73.6 1096.1 1491.2 576.3 1191.8 1212.5 65.3
Nikon Z50 (lossless) 36.3 541.1 716.7 274.9 573.6 591.2 31.5
Nikon D850 (14-bit lossless) 72.7 1102.0 1495.6 574.7 1190.5 1317.3 63.4
Nikon D750 (14-bit lossless) 40.3 597.7 800.2 295.0 620.6 677.3 36.8
Sony A7 IV (lossless) 40.8 869.8 1158.2 220.2 922.5 ❌ 31.9
Sony A7 IV (compressed) 26.3 535.9 814.1 192.5 574.2 662.9 16.8
Sony A7R V (lossless) 70.5 1537.6 2097.5 395.2 1678.6 ❌ 54.4
Sony A1 (compressed) 39.3 797.4 1220.4 283.0 855.0 998.0 23.9
Sony A7 III (compressed) 21.2 393.1 604.4 144.2 410.2 491.4 13.0
Sony A7C II (compressed) 27.5 527.0 817.9 190.4 574.2 662.9 16.9
Sony A9 III (compressed) 20.1 389.4 604.3 140.9 419.1 489.3 13.5
Sony A6700 (lossless) 34.2 735.7 952.2 187.1 766.1 ❌ 27.1
Sony A6400 (compressed) 19.9 396.9 605.1 143.7 422.3 487.9 13.3
Sony ZV-E10 (compressed) 17.7 397.5 604.1 145.1 426.6 430.1 12.6
Fujifilm X-T5 (lossless compressed) 74.4 696.6 1031.5 387.9 1808.9 ❌ 55.1
Fujifilm X-H2 (compressed) 94.8 878.0 1203.3 420.2 2268.9 ❌ 65.1
Fujifilm X-S20 (lossless compressed) 54.0 488.6 711.3 273.1 1217.9 ❌ 41.1
Fujifilm X100VI (lossless compressed) 83.6 904.0 1212.0 385.5 2015.8 ❌ 54.4
Fujifilm X-T4 (uncompressed) 28.3 380.5 603.4 195.8 419.3 434.4 9.09
Fujifilm GFX100 II (lossless compressed) 125.4 1568.7 2451.3 688.2 4106.0 ❌ 99.8
Panasonic Lumix S5II 30.4 476.9 686.2 225.8 622.1 471.1 26.5
Panasonic Lumix GH6 32.9 498.3 717.5 231.5 674.3 508.6 27.4
Panasonic Lumix G9II 32.1 494.5 719.4 229.7 684.6 552.6 25.0
Panasonic Lumix S5 18.2 361.6 569.6 113.0 405.3 482.1 10.8
Panasonic Lumix S1R 38.0 798.7 1205.6 214.5 894.5 896.3 22.9
Panasonic Lumix GH5 17.7 391.0 567.6 97.8 427.4 443.1 12.0
OM System OM-1 Mark II 126.4 677.6 845.5 341.3 694.8 338.4 124.5
OM System OM-5 138.8 639.5 821.2 336.7 654.6 348.1 132.5
Olympus E-M1 Mark III 137.9 628.6 809.8 314.0 634.8 690.3 131.6
Pentax K-3 Mark III 48.2 562.4 772.3 321.4 671.4 764.4 41.3
Pentax K-1 Mark II 60.9 777.3 1094.8 457.9 934.7 1073.3 54.1
Ricoh GR III 34.1 658.9 863.8 343.4 695.1 779.0 28.5
Leica Q2 33.1 926.5 1323.7 259.4 1079.4 1233.0 CPU¹
Leica SL2 33.8 920.6 1301.5 257.6 1072.0 1229.3 CPU¹
Leica M10-R 54.9 1085.2 1456.2 605.4 1162.4 1320.6 45.5
Hasselblad X2D 100C (compressed) 144.2 2315.9 3203.9 ❌ 2614.4 3216.0 118.9
Phase One IQ4 150MP (IIQ L) 133.0 4304.9 5691.9 952.2 4751.1 5987.5 92.5
Sigma fp L 72.1 1565.7 2081.1 364.3 1666.2 1898.8 CPU¹
Apple iPhone 12 Pro (ProRAW) 24.3 591.8 687.6 74.4 594.3 679.3 21.0
Google Pixel 8 Pro 59.1 1445.9 1891.3 303.5 1530.9 1566.9 CPU¹
Google Pixel 6 Pro 16.7 350.0 453.8 85.5 348.8 401.4 CPU¹
Samsung Galaxy S23 Ultra 177.7 2194.2 2507.1 1292.4 2114.0 2609.3 CPU¹
Samsung Galaxy S22 (JPEG XL) 289.3 ❌ ❌ 212.3 ❌ ❌ CPU¹
DJI Phantom 4 Pro 18.9 291.5 474.1 116.9 315.3 394.8 CPU¹
GoPro HERO11 Black (GPR) 60.2 ❌ ❌ ❌ ❌ ❌ CPU¹
GoPro HERO10 Black (GPR) 53.9 ❌ ❌ ❌ ❌ ❌ CPU¹

FlashRaw GPU = --backend gpu --output rgba8: CPU decode or GPU unpack, GPU develop, frame downloaded to host memory. ¹ FlashRaw's GPU pipeline does not handle this file's processing yet and ran it on the CPU.

Popular cameras (Linux x86-64)

Times in ms, all cores, bold = fastest. One-thread tables: result file.

Decode

Camera Format · compression MP FlashRaw LibRaw RawSpeed rawler rawpy dcraw ²
Canon EOS R5 (RAW) CR3 · canon-crx 46.7 57.2 57.9 ❌ 181.1 59.8 ❌
Canon EOS R5 (C-RAW) CR3 · canon-crx 46.7 71.9 68.5 ❌ 150.7 67.4 ❌
Canon EOS R5 Mark II (RAW) CR3 · canon-crx 47.9 57.0 59.0 ❌ 180.9 60.9 ❌
Canon EOS R6 Mark II (RAW) CR3 · canon-crx 25.5 33.1 32.8 ❌ 97.6 33.8 ❌
Canon EOS R6 Mark III (RAW) CR3 · canon-crx 34.0 41.6 42.6 ❌ 126.7 43.8 ❌
Canon EOS R7 (RAW) CR3 · canon-crx 33.7 42.5 45.6 ❌ 127.1 43.7 ❌
Canon EOS R8 (RAW) CR3 · canon-crx 25.5 32.1 33.0 ❌ 101.3 33.8 ❌
Canon EOS R50 (RAW) CR3 · canon-crx 25.1 30.9 31.6 ❌ 94.6 32.8 ❌
Canon EOS 90D (RAW) CR3 · canon-crx 33.7 33.8 41.9 ❌ 98.6 46.2 ❌
Canon EOS 5D Mark IV (RAW) CR2 · lossless-jpeg 31.3 23.9 346.8 68.9 129.8 363.6 403.2
Canon EOS 6D Mark II (RAW) CR2 · lossless-jpeg 27.0 20.5 262.3 82.5 126.7 274.2 305.8
Canon EOS Rebel T7 / 2000D (RAW) CR2 · lossless-jpeg 24.7 24.5 308.9 88.0 131.4 324.8 344.6
Nikon Z8 (14-bit lossless) NEF · nikon-huffman 45.7 46.5 421.0 178.0 203.3 450.6 498.7
Nikon Z9 (14-bit lossless) NEF · nikon-huffman 45.7 51.8 407.3 133.7 172.6 423.4 484.8
Nikon Z6III (lossless) NEF · nikon-huffman 24.5 28.3 228.7 85.4 104.4 243.0 269.9
Nikon Zf (lossless) NEF · nikon-huffman 24.5 27.8 216.9 68.6 90.5 231.2 263.2
Nikon Z7II (lossless) NEF · nikon-huffman 45.7 51.2 403.1 117.1 159.8 422.8 481.2
Nikon Z50 (lossless) NEF · nikon-huffman 20.9 24.3 201.4 72.8 86.9 213.7 237.1
Nikon D850 (14-bit lossless) NEF · nikon-huffman 45.7 51.5 414.7 118.6 159.7 422.3 480.0
Nikon D750 (14-bit lossless) NEF · nikon-huffman 24.3 28.6 204.0 57.6 78.8 218.9 252.8
Sony A7 IV (lossless) ARW · sony-lossless 36.7 14.9 314.2 15.9 36.9 339.5 ❌
Sony A7 IV (compressed) ARW · sony-arw2 33.0 8.02 80.2 10.9 11.9 96.6 172.5
Sony A7R V (lossless) ARW · sony-lossless 64.7 25.6 576.4 31.3 60.8 611.9 ❌
Sony A1 (compressed) ARW · sony-arw2 50.2 10.8 117.8 17.3 18.5 142.2 257.8
Sony A7 III (compressed) ARW · sony-arw2 24.3 5.45 59.1 8.18 8.09 71.4 128.4
Sony A7C II (compressed) ARW · sony-arw2 33.0 7.47 79.3 10.5 11.9 95.4 173.1
Sony A9 III (compressed) ARW · sony-arw2 24.3 5.69 59.7 8.41 8.47 71.7 130.1
Sony A6700 (lossless) ARW · sony-lossless 30.7 13.7 275.5 15.0 34.8 294.2 ❌
Sony A6400 (compressed) ARW · sony-arw2 24.3 5.25 52.7 8.22 8.23 65.7 124.0
Sony ZV-E10 (compressed) ARW · sony-arw2 24.3 5.62 54.0 7.78 8.23 66.5 125.5
Fujifilm X-T5 (lossless compressed) RAF · fuji-compressed 40.9 53.1 155.4 112.1 132.1 153.5 ❌
Fujifilm X-H2 (compressed) RAF · fuji-compressed 40.9 67.0 168.3 ❌ 147.6 183.9 ❌
Fujifilm X-S20 (lossless compressed) RAF · fuji-compressed 26.5 39.4 115.0 83.5 101.1 120.9 ❌
Fujifilm X100VI (lossless compressed) RAF · fuji-compressed 40.9 53.1 156.1 121.5 133.2 157.7 ❌
Fujifilm X-T4 (uncompressed) RAF · uncompressed 26.7 3.35 16.1 4.24 11.8 16.7 114.6
Fujifilm GFX100 II (lossless compressed) RAF · fuji-compressed 103.4 101.6 252.4 202.5 227.5 251.9 ❌
Panasonic Lumix S5II RW2 · panasonic-v8 24.1 24.8 87.5 89.0 81.7 84.7 174.1
Panasonic Lumix GH6 RW2 · panasonic-v8 25.2 28.8 90.9 95.1 85.5 103.1 188.0
Panasonic Lumix G9II RW2 · panasonic-v8 25.1 28.1 92.3 95.5 86.9 89.1 182.0
Panasonic Lumix S5 RW2 · panasonic-v7 24.3 4.80 10.5 4.58 9.47 9.67 171.8
Panasonic Lumix S1R RW2 · panasonic-v6 47.6 12.4 85.7 15.3 17.9 92.9 344.6
Panasonic Lumix GH5 RW2 · panasonic-v4 20.6 6.75 60.5 17.0 9.83 59.9 130.4
OM System OM-1 Mark II ORF · olympus 20.4 64.0 331.8 142.2 187.5 323.8 80.5
OM System OM-5 ORF · olympus 20.5 60.3 275.8 143.7 198.4 266.7 81.0
Olympus E-M1 Mark III ORF · olympus 20.5 57.1 250.7 132.0 166.2 247.5 334.1
Pentax K-3 Mark III PEF · pentax-huffman 26.6 40.2 273.4 57.2 86.6 250.6 269.0
Pentax K-1 Mark II PEF · pentax-huffman 36.6 42.6 316.9 91.6 134.9 356.6 376.8
Ricoh GR III DNG · lossless-jpeg 24.8 25.7 372.0 68.5 104.5 363.5 309.3
Leica Q2 DNG · uncompressed 47.4 8.37 191.7 33.2 9.21 237.0 270.4
Leica SL2 DNG · uncompressed 47.4 8.47 191.0 32.5 9.31 236.8 271.8
Leica M10-R DNG · lossless-jpeg 41.0 43.3 471.7 150.6 192.3 553.3 505.3
Hasselblad X2D 100C (compressed) FFF · hasselblad 105.3 91.7 849.8 490.0 ❌ 1043.7 1203.7
Phase One IQ4 150MP (IIQ L) IIQ · phaseone 154.0 55.7 666.5 ❌ 356.8 960.3 1701.1
Sigma fp L DNG · lossless-jpeg 60.5 17.7 653.5 23.7 53.6 714.9 747.3
Apple iPhone 12 Pro (ProRAW) DNG · lossless-jpeg 12.2 19.6 316.5 ❌ 25.5 350.4 325.5
Google Pixel 8 Pro DNG · lossless-jpeg 50.1 13.8 520.7 18.8 42.5 556.8 645.2
Google Pixel 6 Pro DNG · lossless-jpeg 12.5 3.32 130.2 4.09 12.3 142.8 160.7
Samsung Galaxy S23 Ultra DNG · lossless-jpeg 49.9 102.5 1274.9 380.9 485.1 1408.4 1293.7
Samsung Galaxy S22 (JPEG XL) DNG · jpeg-xl 12.0 236.6 ❌ ❌ 156.4 ❌ ❌
DJI Phantom 4 Pro DNG · uncompressed 20.0 2.63 10.7 3.30 4.85 25.6 79.2
GoPro HERO11 Black (GPR) DNG · gopro-vc5 27.1 48.3 ❌ 27.9 ❌ ❌ ❌
GoPro HERO10 Black (GPR) DNG · gopro-vc5 23.3 48.4 ❌ 23.4 ❌ ❌ ❌

RAW → sRGB

Camera FlashRaw LibRaw LibRaw + Little CMS rawler rawpy dcraw ²
Canon EOS R5 (RAW) 101.2 481.1 1080.1 640.5 529.9 ❌
Canon EOS R5 (C-RAW) 115.0 492.0 1087.4 611.6 539.3 ❌
Canon EOS R5 Mark II (RAW) 99.9 484.3 1081.5 646.8 531.6 ❌
Canon EOS R6 Mark II (RAW) 55.6 264.0 579.4 346.3 289.4 ❌
Canon EOS R6 Mark III (RAW) 72.8 349.8 788.4 462.2 382.6 ❌
Canon EOS R7 (RAW) 72.7 351.1 790.5 463.7 386.3 ❌
Canon EOS R8 (RAW) 55.8 262.2 580.4 344.7 288.8 ❌
Canon EOS R50 (RAW) 54.7 260.0 585.2 342.7 287.2 ❌
Canon EOS 90D (RAW) 69.9 353.4 790.8 433.1 385.8 ❌
Canon EOS 5D Mark IV (RAW) 55.0 633.7 1072.1 446.6 688.4 663.9
Canon EOS 6D Mark II (RAW) 45.5 514.9 869.8 395.8 552.8 530.1
Canon EOS Rebel T7 / 2000D (RAW) 47.6 537.0 856.0 379.8 585.8 553.2
Nikon Z8 (14-bit lossless) 91.0 856.4 1457.4 662.7 922.8 806.9
Nikon Z9 (14-bit lossless) 95.2 838.5 1438.0 631.9 904.3 791.8
Nikon Z6III (lossless) 52.3 464.5 785.8 356.7 502.7 435.4
Nikon Zf (lossless) 53.0 450.1 772.5 343.2 488.6 428.8
Nikon Z7II (lossless) 95.0 835.5 1431.8 627.7 902.9 792.1
Nikon Z50 (lossless) 44.3 402.1 696.6 302.6 434.0 378.8
Nikon D850 (14-bit lossless) 94.5 835.0 1445.6 619.9 896.5 870.5
Nikon D750 (14-bit lossless) 53.2 434.5 748.2 325.4 473.4 460.9
Sony A7 IV (lossless) 47.3 628.4 1073.4 372.3 692.5 ❌
Sony A7 IV (compressed) 40.8 393.7 834.7 344.8 443.0 457.4
Sony A7R V (lossless) 84.7 1151.4 1958.7 671.3 1261.4 ❌
Sony A1 (compressed) 60.6 592.8 1267.4 519.3 667.6 689.7
Sony A7 III (compressed) 31.2 291.1 609.7 258.0 330.7 338.1
Sony A7C II (compressed) 40.8 391.0 834.0 350.3 442.3 455.0
Sony A9 III (compressed) 30.1 289.9 613.1 257.1 330.4 337.2
Sony A6700 (lossless) 39.4 525.2 880.9 301.1 573.6 ❌
Sony A6400 (compressed) 29.7 284.1 601.4 258.0 320.6 332.4
Sony ZV-E10 (compressed) 26.4 285.1 602.4 257.5 323.4 289.3
Fujifilm X-T5 (lossless compressed) 115.6 539.0 1089.3 610.4 573.2 ❌
Fujifilm X-H2 (compressed) 144.6 581.7 1112.9 623.9 636.9 ❌
Fujifilm X-S20 (lossless compressed) 78.9 370.9 723.9 413.6 398.7 ❌
Fujifilm X100VI (lossless compressed) 130.8 567.2 1098.7 613.4 607.8 ❌
Fujifilm X-T4 (uncompressed) 43.6 266.9 618.7 323.6 293.1 312.3
Fujifilm GFX100 II (lossless compressed) 199.5 1208.7 2583.7 1241.5 1305.5 ❌
Panasonic Lumix S5II 52.2 352.0 664.1 299.3 388.9 342.4
Panasonic Lumix GH6 56.2 366.9 728.8 312.1 404.8 360.4
Panasonic Lumix G9II 51.1 364.3 700.8 312.6 401.9 401.2
Panasonic Lumix S5 29.3 273.5 589.0 225.2 312.1 339.4
Panasonic Lumix S1R 59.2 595.0 1210.6 430.9 672.6 671.3
Panasonic Lumix GH5 26.6 280.0 638.6 193.0 317.9 305.7
OM System OM-1 Mark II 86.2 527.6 801.4 396.1 539.2 213.7
OM System OM-5 89.5 470.5 769.1 398.8 489.6 212.6
Olympus E-M1 Mark III 78.1 461.6 775.8 418.6 485.0 507.5
Pentax K-3 Mark III 64.1 475.4 879.8 370.1 663.0 498.3
Pentax K-1 Mark II 88.5 659.5 1194.2 514.8 738.9 691.4
Ricoh GR III 86.6 730.6 1235.3 376.4 584.3 517.1
Leica Q2 53.9 640.7 1268.2 490.5 735.2 677.3
Leica SL2 54.3 639.4 1245.3 481.8 733.1 679.1
Leica M10-R 85.1 950.1 1377.6 605.2 902.7 857.2
Hasselblad X2D 100C (compressed) 193.6 1821.3 4040.5 ❌ 2097.8 1916.8
Phase One IQ4 150MP (IIQ L) 227.2 2442.2 4703.7 1670.4 3177.4 3009.8
Sigma fp L 96.4 1192.7 2083.7 823.5 1323.3 1270.6
Apple iPhone 12 Pro (ProRAW) 35.7 382.5 535.8 121.9 427.8 458.4
Google Pixel 8 Pro 69.6 1088.7 1766.2 805.6 1217.3 1065.2
Google Pixel 6 Pro 18.2 250.8 419.4 143.1 283.3 267.7
Samsung Galaxy S23 Ultra 202.4 1490.4 2079.0 850.6 1709.5 1836.7
Samsung Galaxy S22 (JPEG XL) 293.2 ❌ ❌ 251.4 ❌ ❌
DJI Phantom 4 Pro 33.8 201.8 473.8 214.2 237.2 250.9
GoPro HERO11 Black (GPR) 77.2 ❌ ❌ ❌ ❌ ❌
GoPro HERO10 Black (GPR) 67.2 ❌ ❌ ❌ ❌ ❌

Memory

Bar chart: peak memory relative to FlashRaw

Peak resident memory of the benchmark process (input file included), as a ratio to FlashRaw's on the same file (> 1 = FlashRaw needs less), median over the popular cameras, and the number of files where FlashRaw needs less:

Comparison threads 🐧 ratio 🐧 less 🍎 ratio 🍎 less
decode vs LibRaw all 0.91× 1/59 0.81× 1/59
decode vs LibRaw 1 0.98× 2/59 0.94× 2/59
decode vs RawSpeed all 0.98× 21/49 0.90× 0/49
decode vs RawSpeed 1 1.01× 42/49 0.99× 0/49
decode vs rawler all 1.14× 40/59 1.54× 39/59
decode vs rawler 1 1.49× 57/59 1.61× 57/59
decode vs rawpy all 1.05× 30/59 0.98× 28/59
decode vs rawpy 1 1.17× 45/59 1.11× 45/59
decode vs dcraw all 2.02× 38/42 1.94× 33/42
decode vs dcraw 1 2.86× 40/42 2.75× 40/42
RAW→RGB vs LibRaw all 1.88× 59/59 1.48× 54/59
RAW→RGB vs LibRaw 1 1.92× 59/59 1.88× 59/59
RAW→sRGB vs LibRaw + Little CMS all 2.78× 59/59 2.21× 59/59
RAW→sRGB vs LibRaw + Little CMS 1 2.86× 59/59 2.79× 59/59
RAW→RGB vs rawler all 4.14× 59/59 4.55× 59/59
RAW→RGB vs rawler 1 4.32× 59/59 5.14× 59/59
RAW→RGB vs rawpy all 1.97× 59/59 1.55× 56/59
RAW→RGB vs rawpy 1 2.03× 59/59 1.97× 59/59
RAW→RGB vs dcraw all 1.31× 36/42 0.94× 20/42
RAW→RGB vs dcraw 1 1.34× 40/42 1.32× 40/42

Median peaks in MB:

Median peak RSS (MB), all cores FlashRaw LibRaw RawSpeed rawler rawpy dcraw LibRaw + Little CMS
🐧 decode 183 104 109 178 129 247 –
🐧 RAW → sRGB 274 415 – 848 440 247 619
🍎 decode 174 104 109 196 122 247 –
🍎 RAW → sRGB 303 416 – 1083 435 247 620

Each median is over the files that library decodes, so the columns are not over identical sets; the ratios above are per file.

  • RAW → sRGB: FlashRaw needs less memory than every library pipeline, typically half to two thirds of LibRaw's (and rawpy's) and less than half of the LibRaw + Little CMS pipeline's. It develops in one fused pass straight into the output image, so it never holds full-size intermediates: LibRaw keeps a 4-channel 16-bit working image plus its output, the Little CMS pipeline adds a 16-bit RGB copy and the 8-bit result, and rawler develops in 32-bit floats (4–5× FlashRaw's peak). Only dcraw, a single program with one 4 × 16-bit image buffer and no threads, needs about as little: its peak is a median 1.3× FlashRaw's on Linux and 0.94× on the M1 Ultra (all cores).
  • Decode only: about the same on one thread, more for FlashRaw on all cores. Every decoder holds the file and a 16-bit image. On one thread FlashRaw's peak is within a few percent of LibRaw's and RawSpeed's. With all cores, FlashRaw's parallel decoders of single entropy-coded streams and tiled codecs add per-thread working buffers (for example, each thread decodes its chunk of the stream into its own buffer of differences before the rows are rebuilt; see architecture): LibRaw's peak is a median 0.9× of FlashRaw's on Linux and 0.8× on the M1 Ultra. That is the price of decoding Nikon, lossless-JPEG or Pentax streams 5–10× faster.

Per camera (Apple M1 Ultra, all cores, MB):

Camera MP decode: FlashRaw LibRaw RawSpeed rawler rawpy dcraw ² RGB: FlashRaw LibRaw LibRaw + Little CMS rawler rawpy dcraw ²
Canon EOS R5 (RAW) 46.7 227 134 ❌ 409 152 ❌ 409 604 904 1886 621 ❌
Canon EOS R5 (C-RAW) 46.7 216 127 ❌ 397 141 ❌ 394 596 896 1875 611 ❌
Canon EOS R5 Mark II (RAW) 47.9 247 149 ❌ 427 168 ❌ 427 622 926 1916 642 ❌
Canon EOS R6 Mark II (RAW) 25.5 137 83 ❌ 239 101 ❌ 233 337 499 1039 356 ❌
Canon EOS R6 Mark III (RAW) 34.0 173 104 ❌ 278 122 ❌ 303 442 659 1347 460 ❌
Canon EOS R7 (RAW) 33.7 172 102 ❌ 275 119 ❌ 300 444 661 1342 461 ❌
Canon EOS R8 (RAW) 25.5 136 81 ❌ 239 99 ❌ 233 337 500 1038 354 ❌
Canon EOS R50 (RAW) 25.1 138 85 ❌ 239 103 ❌ 234 337 497 1034 354 ❌
Canon EOS 90D (RAW) 33.7 174 104 ❌ 278 122 ❌ 303 445 664 1347 463 ❌
Canon EOS 5D Mark IV (RAW) 31.3 294 97 100 190 115 294 414 416 620 1126 435 294
Canon EOS 6D Mark II (RAW) 27.0 237 91 95 126 110 254 333 368 544 984 387 254
Canon EOS Rebel T7 / 2000D (RAW) 24.7 266 89 92 170 107 234 363 344 506 911 362 234
Nikon Z8 (14-bit lossless) 45.7 294 149 153 235 169 439 477 630 936 1719 648 439
Nikon Z9 (14-bit lossless) 45.7 304 147 150 228 165 439 476 627 934 1712 646 439
Nikon Z6III (lossless) 24.5 184 83 86 144 102 236 276 340 505 940 358 236
Nikon Zf (lossless) 24.5 176 81 84 141 99 236 258 339 503 937 358 236
Nikon Z7II (lossless) 45.7 293 144 147 224 162 439 478 626 931 1709 643 439
Nikon Z50 (lossless) 20.9 165 74 78 133 93 202 250 294 434 810 312 202
Nikon D850 (14-bit lossless) 45.7 294 144 148 224 164 439 474 626 932 1710 643 439
Nikon D750 (14-bit lossless) 24.3 166 76 79 133 95 235 257 332 495 925 351 235
Sony A7 IV (lossless) 36.7 140 122 126 216 141 ❌ 267 468 688 1323 487 ❌
Sony A7 IV (compressed) 33.0 112 106 111 180 125 318 242 452 674 1251 471 317
Sony A7R V (lossless) 64.7 230 211 215 343 229 ❌ 460 852 1259 2359 869 ❌
Sony A1 (compressed) 50.2 166 160 165 257 180 481 367 687 1022 1883 705 481
Sony A7 III (compressed) 24.3 81 76 80 134 95 234 177 330 493 924 348 234
Sony A7C II (compressed) 33.0 111 105 110 179 124 318 242 453 673 1250 472 318
Sony A9 III (compressed) 24.3 84 79 83 141 97 234 181 333 496 930 352 234
Sony A6700 (lossless) 30.7 127 105 109 196 124 ❌ 224 378 553 1083 397 ❌
Sony A6400 (compressed) 24.3 80 74 79 135 94 234 177 330 493 922 348 234
Sony ZV-E10 (compressed) 24.3 80 74 79 135 94 234 162 330 494 923 349 234
Fujifilm X-T5 (lossless compressed) 40.9 130 126 128 197 143 ❌ 287 547 816 1508 565 ❌
Fujifilm X-H2 (compressed) 40.9 117 113 ❌ 175 130 ❌ 275 534 803 1487 551 ❌
Fujifilm X-S20 (lossless compressed) 26.5 93 89 91 150 107 ❌ 197 363 539 1000 381 ❌
Fujifilm X100VI (lossless compressed) 40.9 135 132 134 208 149 ❌ 293 554 822 1517 570 ❌
Fujifilm X-T4 (uncompressed) 26.7 115 110 113 197 129 252 220 384 560 998 403 252
Fujifilm GFX100 II (lossless compressed) 103.4 314 311 312 455 327 ❌ 713 1382 2062 3779 1398 ❌
Panasonic Lumix S5II 24.1 271 89 89 151 107 232 370 343 505 931 359 232
Panasonic Lumix GH6 25.2 290 100 98 169 115 243 386 364 532 982 380 243
Panasonic Lumix G9II 25.1 262 96 95 162 113 242 360 359 528 977 377 242
Panasonic Lumix S5 24.3 97 92 96 169 111 234 195 347 509 951 367 234
Panasonic Lumix S1R 47.6 167 161 166 264 181 453 355 658 973 1787 677 453
Panasonic Lumix GH5 20.6 72 66 71 125 86 197 153 281 418 783 299 197
OM System OM-1 Mark II 20.4 71 65 68 120 84 197 154 281 418 782 298 197
OM System OM-5 20.5 70 63 65 114 82 198 150 278 416 781 297 198
Olympus E-M1 Mark III 20.5 67 60 64 112 80 198 149 277 415 777 295 198
Pentax K-3 Mark III 26.6 182 85 89 121 105 257 285 359 532 968 376 257
Pentax K-1 Mark II 36.6 260 120 123 153 138 352 405 502 746 1337 521 352
Ricoh GR III 24.8 257 84 87 166 102 235 352 338 501 911 357 235
Leica Q2 47.4 184 177 181 213 197 454 374 675 992 1748 693 454
Leica SL2 47.4 186 180 183 216 199 454 375 675 991 1749 693 454
Leica M10-R 41.0 407 134 138 248 153 394 564 567 841 1500 585 394
Hasselblad X2D 100C (compressed) 105.3 596 359 362 ❌ 377 1007 1066 1431 2113 ❌ 1448 1007
Phase One IQ4 150MP (IIQ L) 154.0 466 458 ❌ 655 477 1451 1059 2342 3352 5586 2360 1451
Sigma fp L 60.5 218 208 213 375 227 580 457 844 1249 2224 862 580
Apple iPhone 12 Pro (ProRAW) 12.2 124 126 ❌ 191 144 96 173 254 337 716 272 96
Google Pixel 8 Pro 50.1 153 144 148 291 163 481 348 671 1007 1822 689 481
Google Pixel 6 Pro 12.5 49 40 44 111 59 122 98 174 257 494 190 122
Samsung Galaxy S23 Ultra 49.9 728 463 371 687 481 384 744 986 1320 2402 1006 384
Samsung Galaxy S22 (JPEG XL) 12.0 655 ❌ ❌ 506 ❌ ❌ 695 ❌ ❌ 991 ❌ ❌
DJI Phantom 4 Pro 20.0 86 80 84 115 99 193 210 292 425 763 310 193
GoPro HERO11 Black (GPR) 27.1 340 ❌ 287 ❌ ❌ ❌ 440 ❌ ❌ ❌ ❌ ❌
GoPro HERO10 Black (GPR) 23.3 290 ❌ 247 ❌ ❌ ❌ 375 ❌ ❌ ❌ ❌ ❌

LibRaw + Little CMS pipeline

The colour-managed pipeline split into its stages, all cores. Medians over the popular cameras: on the M1 Ultra the pipeline takes a median 846 ms (LibRaw decode 194, develop 348, Little CMS 282); on the Ryzen 856 ms (194, 280 and 430). On one thread the Little CMS stage dominates (84–85 % of the pipeline): a 16-bit → 8-bit matrix-shaper transform through Little CMS's generic optimised pipeline costs 90–130 ns per pixel, which is why it was run over row blocks on all cores here (Little CMS's own multithreading and SIMD plugins are GPL-3 or commercial, and were not used). FlashRaw needs no separate colour stage: its single pass applies the white balance, the camera-to-sRGB matrix and the sRGB curve per pixel while demosaicing.

Per camera (Apple M1 Ultra, ms):

decode = LibRaw open_buffer + unpack; develop = dcraw_process to 16-bit linear ProPhoto + dcraw_make_mem_image; Little CMS = profiles, transform and conversion to 8-bit sRGB RGBA; FlashRaw = RAW → RGBA8 in one pass.

Camera LibRaw decode LibRaw develop Little CMS total FlashRaw speedup
Canon EOS R5 (RAW) 136.6 561.0 406.6 1104.5 109.3 10.1×
Canon EOS R5 (C-RAW) 140.9 566.0 409.2 1116.2 129.4 8.63×
Canon EOS R5 Mark II (RAW) 142.4 566.8 413.0 1122.9 111.7 10.0×
Canon EOS R6 Mark II (RAW) 78.3 318.3 222.7 617.0 61.0 10.1×
Canon EOS R6 Mark III (RAW) 100.6 416.8 297.2 815.7 80.4 10.2×
Canon EOS R7 (RAW) 101.2 417.9 298.6 818.9 80.3 10.2×
Canon EOS R8 (RAW) 76.6 316.2 224.5 612.5 60.8 10.1×
Canon EOS R50 (RAW) 75.8 313.1 224.7 610.7 60.3 10.1×
Canon EOS 90D (RAW) 101.7 418.9 301.8 822.0 52.7 15.6×
Canon EOS 5D Mark IV (RAW) 430.2 387.7 282.5 1100.4 39.8 27.7×
Canon EOS 6D Mark II (RAW) 388.7 341.4 242.3 968.6 33.7 28.7×
Canon EOS Rebel T7 / 2000D (RAW) 394.9 316.0 216.4 931.9 34.3 27.2×
Nikon Z8 (14-bit lossless) 545.7 583.8 411.3 1541.7 72.4 21.3×
Nikon Z9 (14-bit lossless) 495.2 571.7 414.1 1484.4 72.4 20.5×
Nikon Z6III (lossless) 282.0 312.2 227.0 824.4 41.5 19.9×
Nikon Zf (lossless) 269.2 322.3 230.1 819.7 41.9 19.5×
Nikon Z7II (lossless) 493.9 579.2 417.9 1491.2 73.6 20.3×
Nikon Z50 (lossless) 248.4 276.5 193.1 716.7 36.3 19.7×
Nikon D850 (14-bit lossless) 488.1 586.7 423.9 1495.6 72.7 20.6×
Nikon D750 (14-bit lossless) 259.5 322.2 218.9 800.2 40.3 19.9×
Sony A7 IV (lossless) 425.3 423.4 305.7 1158.2 40.8 28.4×
Sony A7 IV (compressed) 86.8 424.7 299.6 814.1 26.3 30.9×
Sony A7R V (lossless) 773.3 764.8 558.6 2097.5 70.5 29.7×
Sony A1 (compressed) 130.7 633.2 457.9 1220.4 39.3 31.1×
Sony A7 III (compressed) 63.7 319.4 223.3 604.4 21.2 28.6×
Sony A7C II (compressed) 88.9 418.4 307.7 817.9 27.5 29.7×
Sony A9 III (compressed) 63.4 318.9 222.7 604.3 20.1 30.0×
Sony A6700 (lossless) 375.4 334.6 244.3 952.2 34.2 27.9×
Sony A6400 (compressed) 64.2 314.3 224.8 605.1 19.9 30.5×
Sony ZV-E10 (compressed) 64.2 319.0 222.6 604.1 17.7 34.1×
Fujifilm X-T5 (lossless compressed) 151.5 507.7 368.7 1031.5 74.4 13.9×
Fujifilm X-H2 (compressed) 176.5 667.5 361.1 1203.3 94.8 12.7×
Fujifilm X-S20 (lossless compressed) 128.1 341.2 245.8 711.3 54.0 13.2×
Fujifilm X100VI (lossless compressed) 154.5 695.9 363.8 1212.0 83.6 14.5×
Fujifilm X-T4 (uncompressed) 19.9 341.5 243.0 603.4 28.3 21.3×
Fujifilm GFX100 II (lossless compressed) 277.1 1258.5 919.7 2451.3 125.4 19.5×
Panasonic Lumix S5II 130.4 335.4 220.6 686.2 30.4 22.6×
Panasonic Lumix GH6 133.7 348.0 232.7 717.5 32.9 21.8×
Panasonic Lumix G9II 138.3 348.0 235.7 719.4 32.1 22.4×
Panasonic Lumix S5 9.57 335.6 224.7 569.6 18.2 31.3×
Panasonic Lumix S1R 124.2 652.0 429.2 1205.6 38.0 31.7×
Panasonic Lumix GH5 91.5 282.4 194.4 567.6 17.7 32.1×
OM System OM-1 Mark II 389.5 273.8 187.8 845.5 126.4 6.69×
OM System OM-5 355.0 267.0 193.6 821.2 138.8 5.92×
Olympus E-M1 Mark III 344.9 271.2 192.2 809.8 137.9 5.87×
Pentax K-3 Mark III 193.5 336.4 241.0 772.3 48.2 16.0×
Pentax K-1 Mark II 291.0 464.9 336.9 1094.8 60.9 18.0×
Ricoh GR III 322.1 314.6 225.5 863.8 34.1 25.3×
Leica Q2 285.9 594.7 435.7 1323.7 33.1 40.0×
Leica SL2 288.0 593.7 417.4 1301.5 33.8 38.5×
Leica M10-R 556.7 524.1 375.6 1456.2 54.9 26.5×
Hasselblad X2D 100C (compressed) 1001.1 1264.5 930.3 3203.9 144.2 22.2×
Phase One IQ4 150MP (IIQ L) 1036.0 3305.0 1347.6 5691.9 133.0 42.8×
Sigma fp L 765.9 761.8 554.5 2081.1 72.1 28.9×
Apple iPhone 12 Pro (ProRAW) 466.7 110.8 110.2 687.6 24.3 28.3×
Google Pixel 8 Pro 652.0 783.8 457.3 1891.3 59.1 32.0×
Google Pixel 6 Pro 163.9 172.4 117.5 453.8 16.7 27.2×
Samsung Galaxy S23 Ultra 1899.3 252.7 353.6 2507.1 177.7 14.1×
DJI Phantom 4 Pro 19.3 271.0 185.1 474.1 18.9 25.0×

Per format

Geometric mean speedup of FlashRaw per (format, codec), all cores, Apple M1 Ultra: LibRaw and RawSpeed over the 159 files of both lists, the other libraries over the popular cameras. ❌ = the other library decoded none of these files; – = not measured. Linux: result file.

Format Compression files decode vs LibRaw decode vs RawSpeed decode vs rawler decode vs rawpy decode vs dcraw RGB vs LibRaw RGB vs LibRaw + LCMS RGB vs rawler RGB vs rawpy RGB vs dcraw
3FR hasselblad 2 12.2× 5.62× – – – 16.1× – – – –
3FR uncompressed 2 0.81× 0.52× – – – 18.7× – – – –
ARW other 2 11.2× ❌ – – – 10.2× – – – –
ARW sony-arw1 2 6.21× 3.41× – – – 11.8× – – – –
ARW sony-arw2 9 12.7× 1.40× 1.44× 14.4× 45.3× 18.3× 30.7× 7.23× 21.4× 24.4×
ARW sony-lossless 4 17.2× 0.70× 2.10× 22.5× ❌ 16.1× 28.7× 5.49× 22.9× ❌
ARW uncompressed 2 8.59× 0.43× – – – 12.9× – – – –
CR2 canon-sraw 2 16.1× 3.94× – – – 14.5× – – – –
CR2 lossless-jpeg 5 20.6× 5.88× 12.6× 20.9× 29.1× 20.7× 27.9× 10.9× 22.2× 24.6×
CR3 canon-crx 11 1.74× ❌ 2.76× 5.48× ❌ 6.81× 10.4× 4.39× 10.2× ❌
CRW canon-crw 2 2.71× 2.00× – – – 5.17× – – – –
CRW uncompressed 1 7.72× ❌ – – – 12.7× – – – –
DCR kodak-65000 2 2.65× 3.54× – – – 7.81× – – – –
DCS lossless-jpeg 1 11.2× 2.95× – – – 15.5× – – – –
DCS uncompressed 2 4.66× 34.9× – – – 20.8× – – – –
DNG deflate 2 13.7× 1.03× – – – 17.7× – – – –
DNG gopro-vc5 2 ❌ 0.48× ❌ ❌ ❌ ❌ ❌ ❌ ❌ ❌
DNG jpeg-xl 1 ❌ ❌ 0.63× ❌ ❌ ❌ ❌ 0.73× ❌ ❌
DNG lossless-jpeg 9 21.5× 2.63× 3.67× 22.5× 32.3× 19.4× 25.4× 6.14× 20.6× 23.4×
DNG uncompressed 5 28.5× 4.45× 1.13× 48.6× 148.2× 20.3× 33.8× 7.17× 25.8× 30.5×
ERF uncompressed 2 8.74× 1.44× – – – 15.4× – – – –
FFF hasselblad 2 12.8× 5.87× ❌ 14.5× 28.9× 16.0× 22.2× ❌ 18.1× 22.3×
IIQ phaseone 3 22.5× 28.4× 8.10× 25.2× 98.6× 31.3× 42.8× 7.16× 35.7× 45.0×
IIQ phaseone-s 2 14.2× ❌ – – – 28.0× – – – –
IIQ uncompressed 1 0.77× ❌ – – – 37.2× – – – –
KDC kodak-dc120 1 1.97× ❌ – – – 4.49× – – – –
KDC other 1 1.48× ❌ – – – 2.89× – – – –
KDC uncompressed 2 52.7× 10.8× – – – 22.4× – – – –
MEF uncompressed 1 55.5× 11.1× – – – 26.8× – – – –
MOS leaf 2 12.6× ❌ – – – 15.9× – – – –
MOS uncompressed 1 11.9× 11.3× – – – 21.9× – – – –
MRW minolta-packed 2 56.0× 11.6× – – – 22.4× – – – –
MRW uncompressed 2 9.00× 8.29× – – – 19.5× – – – –
NEF nikon-huffman 9 10.8× 3.99× 7.19× 11.4× 17.4× 15.0× 20.2× 7.77× 16.2× 16.9×
NEF uncompressed 2 6.19× 2.75× – – – 20.8× – – – –
NRW uncompressed 2 2.99× 1.34× – – – 18.4× – – – –
ORF olympus 5 2.93× 1.26× 1.77× 2.95× 1.76× 4.79× 6.15× 2.46× 4.92× 3.23×
ORF uncompressed 2 16.3× 2.69× – – – 20.2× – – – –
PEF pentax-huffman 4 7.36× 3.16× 6.28× 10.2× 15.9× 12.4× 17.0× 7.08× 14.6× 16.7×
PEF uncompressed 2 18.6× 8.09× – – – 21.4× – – – –
RAF fuji-compressed 7 4.26× 1.22× 2.59× 32.9× ❌ 10.2× 14.6× 4.95× 25.3× ❌
RAF uncompressed 3 4.48× 0.78× 2.03× 10.4× 100.0× 15.2× 21.3× 6.91× 14.8× 15.3×
RAW uncompressed 3 7.81× 12.5× – – – 16.0× – – – –
RW2 panasonic-v4 3 18.7× 1.02× 1.50× 19.5× 43.6× 21.6× 32.1× 5.53× 24.2× 25.0×
RW2 panasonic-v5 2 6.23× 0.63× – – – 18.8× – – – –
RW2 panasonic-v6 2 13.0× 0.94× 1.27× 14.3× 54.9× 20.6× 31.7× 5.64× 23.5× 23.6×
RW2 panasonic-v7 3 3.62× 0.25× 1.45× 3.28× 92.3× 20.3× 31.3× 6.21× 22.3× 26.5×
RW2 panasonic-v8 5 8.09× 3.70× 7.58× 16.1× 15.6× 15.3× 22.3× 7.20× 20.7× 16.0×
RW2 uncompressed 2 8.52× 0.73× – – – 19.2× – – – –
SRW samsung-v0 2 10.5× 3.47× – – – 15.8× – – – –
SRW samsung-v1 2 2.67× 1.10× – – – 5.12× – – – –
SRW samsung-v2 2 1.75× 0.92× – – – 3.82× – – – –
SRW uncompressed 2 49.1× 9.60× – – – 23.6× – – – –
X3F sigma-x3f 2 6.27× ❌ – – – 7.07× – – – –
other kodak-ycbcr 1 8.75× ❌ – – – 9.59× – – – –
other uncompressed 2 8.54× 11.4× – – – 19.6× – – – –

Where FlashRaw is not the fastest

Honest list, from the result files (popular cameras and the 159-file set):

  • RAW → sRGB with all cores: never, against any library, on either machine, except the one JPEG XL DNG (Samsung Galaxy S22), where rawler, whose JPEG XL decoder is faster, wins by 1.2× (Linux) and 1.4× (macOS). On one thread LibRaw is faster on 5 (Linux) and 4 (macOS) of 156 files, mostly large uncompressed files and old Kodak formats (up to 2.8× on a Kodak DC120), and dcraw on the two OM System ORFs (1.04–1.14×).
  • Decode with all cores, Linux: 16 losses in 442 comparisons (file × other library).
    • Canon CR3: close to LibRaw's OpenMP build, which also decodes CR3 in parallel; LibRaw is 1–6 % faster on three CR3 files (rawpy, which is LibRaw, is 7 % faster on one) and slower on the others (on the M1 Ultra FlashRaw is 1.7× faster on all of them). LibRaw is also faster on two small uncompressed files (Nikon Coolpix A1000, GITUP: 1.3×).
    • RawSpeed, 9 of 126 files: GoPro GPR (1.7–2.1×), the Sony pixel-shift ARQ (3.1×), small uncompressed files that take 1–4 ms (Nikon Coolpix A1000, GITUP, Epson R-D1, Fujifilm X-E2: up to 1.5×), a Canon EOS 500D CR2 and the Panasonic Lumix S5 (v7).
    • rawler: the JPEG XL DNG (1.5×).
  • Decode with all cores, Apple M1 Ultra: 38 losses in 442 comparisons, mostly against RawSpeed (30 of 126 files):
    • Memory-bound uncompressed and packed data (3FR, RAF, NEF, NRW, ARW, RW2, Panasonic v5–v7): RawSpeed 1.1–2× faster, and 4× on Panasonic v7 (Lumix S5) and the Sony pixel-shift ARQ. These take 1–10 ms in every library; FlashRaw caps such copy kernels at a few threads, which is right on the Ryzen but leaves bandwidth unused on the M1 Ultra. LibRaw is also faster on 7 of these files (1.1–1.4×).
    • Sony lossless ARW: RawSpeed 1.3–1.6× faster. Tiled lossless-JPEG DNG from phones and the Sigma fp L: RawSpeed 1.2–1.3× faster. GoPro GPR: RawSpeed 2× faster. Samsung NX1/NX500 (V2): RawSpeed 1.1× faster (a strictly sequential codec).
    • rawler: the JPEG XL DNG (1.6×).
  • One thread (decode): FlashRaw beats LibRaw on 143 (Linux) and 139 (macOS) of 156 files, rawler on 57 and 56 of 59, rawpy on 58 and 57 of 59, dcraw on 40 of 42 (both machines; the exceptions are the two OM System ORFs). Against RawSpeed it is faster on 96 of 126 files on Linux, but only 40 of 126 on the M1 Ultra (13 of the 49 popular cameras), where RawSpeed's single-threaded decoders are 1.2× faster on the popular cameras (geometric mean). FlashRaw's lead with all cores comes from parallelising the codecs the others run on one thread.
  • Memory: a plain decode on all cores needs more memory with FlashRaw than with LibRaw or RawSpeed (see Memory); for RAW → sRGB on all cores dcraw's peak is about the same as FlashRaw's (lower on 22 of 42 files on the M1 Ultra).
  • Formats the others do not decode at all are not in these counts: Canon CR3 (RawSpeed, dcraw), Phase One IIQ S and 150 MP, Sigma X3F (RawSpeed), GoPro GPR (LibRaw, rawpy, rawler, dcraw), JPEG XL DNG (LibRaw, rawpy, RawSpeed, dcraw), iPhone ProRAW (RawSpeed), current Fujifilm compressed RAF (dcraw), Hasselblad X2D compressed FFF (rawler).

FlashRaw alone

Results of tools/bench/run.py (all formats of the benchmark set, both modes): linux-x86_64.md, macos-arm64.md. Median throughput over the set:

decode RAW → RGBA8
Linux, Ryzen 7 9700X (16 threads, CPU) 2,061 MP/s 583 MP/s
macOS, M1 Ultra (20 threads, Backend::Auto: GPU develop where it wins) 3,148 MP/s 1,122 MP/s

Uncompressed and packed formats run at memory bandwidth; entropy-coded formats are bound by their entropy decoder (parallel over tiles, strips or blocks where the format has them, and by speculation over single streams otherwise — see architecture.md). Single sequential streams are decoded in parallel by speculation (lossless JPEG, Hasselblad, Nikon, Pentax, Panasonic v8 strips, Sony ARW1). Samsung V2 cannot be: rows start wherever the previous one ended and are predicted from earlier rows, so it runs on one thread with a staged block decoder.

Kernels (criterion, cargo bench -p flashraw --bench kernels)

Kernel Throughput
12-bit MSB unpack, scalar reference 1.05 G samples/s
12-bit MSB unpack, SSE4.1 8.8 G samples/s
16-bit LE unpack (fast / SIMD) 35–37 G samples/s
Lossless JPEG decode, 2 components (synthetic) 617 M samples/s

Thread scaling and memory-bound kernels

Unpack/copy kernels are memory-bound: on the Ryzen they peak at ~4 threads and get 1.7–1.9× slower at 16 threads (memory-controller and page-fault contention), e.g. a 24.6 MP 14-bit DNG unpacks in 7.8 ms (1 thread), 3.0 ms (4), 5.2 ms (16). FlashRaw caps such kernels with par::memory_bound_workers() (threads/4, clamped to 2–8) while entropy decoders and the develop pass use all cores. Large output buffers are advised as transparent huge pages (MADV_HUGEPAGE), cutting page faults ~500×.

A single-threaded decode (threads: 1, --threads 1) runs on the calling thread, exactly like RAYON_NUM_THREADS=1. (It used to run on a dedicated one-worker pool, whose thread macOS could schedule on an efficiency core: about 3× slower on the M1 Ultra in one benchmark round.) Explicit pools (threads: n) run their workers with the user-initiated quality of service on Apple platforms.

GPU (Apple M1 Ultra, Metal)

The "FlashRaw GPU" column of the macOS tables is end to end through the CLI (per-process timing, frame downloaded to host memory). The GPU develop pass replaces the CPU one, so it helps most where development dominates: on the popular cameras RAW → RGB is 1.3× faster with the GPU than with 20 CPU threads (geometric mean over the 52 files where it ran; up to 3.1× for packed and uncompressed data) and 17.1× faster than LibRaw. Some files fall back to the CPU (marked CPU¹ in the result file). Resident output (no download) and per-kernel timings are in gpu.md (cargo bench -p flashraw-gpu --bench gpu_pipeline, median of 7):

File Case CPU GPU (download) GPU (resident)
Leica Q2 47 MP, packed 14-bit raw u16 unpack 2.3 ms 6.4 ms 3.4 ms
Leica Q2 RAW → RGBA8 34.5 ms 13.9 ms 8.5 ms
Leica Q2 RAW → RGBA8 half-size 23.7 ms 6.9 ms 5.3 ms
Nokia 1020 38 MP, packed 10-bit RAW → RGBA8 27.8 ms 11.4 ms 6.6 ms
Leica CL 24.6 MP RAW → RGBA8 19.2 ms 7.7 ms 4.8 ms
Pentax 645Z 52 MP, LJPEG RAW → RGBA8 278.0 ms 261.1 ms 258.1 ms

The GPU wins the development pipeline (2.5–4× resident) and loses raw unpacking (the CPU is at memory bandwidth); for entropy-coded files the CPU decode dominates. Backend::Auto picks the faster side. No discrete GPU was available for measurement.

WASM

From wasm.md: single-threaded WASM runs lossless JPEG at 1.2–1.6× native single-thread time and the RAW → RGBA8 pipeline at 1.2–1.4×; the threaded build lands at 1.3–2.2× native all-core time on 46–103 MP files (M1 Ultra), and WebGPU develops a 42 MP uncompressed ARW to a resident texture in 15 ms. SIMD128 halves packed unpack time vs the non-SIMD build.

Bundle size with all decoders (wasm-opt -O3), from wasm.md, which has every variant (threads, WebGPU, JPEG XL):

Binary Raw gzip -9 brotli -11
flashraw_simd.wasm 2897 KiB 1006 KiB 756 KiB
flashraw.wasm (fallback) 2847 KiB 986 KiB 738 KiB

Performance regression detection

tools/bench/run.py ... --baseline <old.jsonl> --tolerance 15 prints per-file before/after and fails on slowdowns beyond the tolerance (used by perf.yml, which benchmarks the PR and main on the same runner).