XYB-jpeg: jpeg-specific clustering and more progression. #1875
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Benchmark before: ``` Encoding kPixels Bytes BPP E MP/s D MP/s Max norm pnorm BPP*pnorm Bugs ------------------------------------------------------------------------------------------------------------------ jpeg:libjxl:d1.0 13270 3197015 1.9272988 5.469 29.993 2.19093180 0.70477101 1.358304300765 0 ``` Benchmark after: ``` Encoding kPixels Bytes BPP E MP/s D MP/s Max norm pnorm BPP*pnorm Bugs ------------------------------------------------------------------------------------------------------------------ jpeg:libjxl:d1.0 13270 3163678 1.9072018 4.031 25.939 2.19093180 0.70477101 1.344140529099 0 ```
jyrkialakuijala
approved these changes
Nov 7, 2022
Contributor
jyrkialakuijala
left a comment
There was a problem hiding this comment.
slowdown is substantial here on both encoding and decoding, it is unclear if this is a good compromise -- let's keep this in mind as an option to reverse later if we need more speed
Unused-Account
added a commit
to Unused-Account/libjxl
that referenced
this pull request
Dec 7, 2022
* Add first version of streaming JPEG decoder. Benchmark before: ``` Encoding kPixels Bytes BPP E MP/s D MP/s Max norm pnorm BPP*pnorm Bugs ------------------------------------------------------------------------------------------------------------------------ jpeg:q90 13270 4838710 2.9169834 63.705 161.410 2.40070152 0.68663652 2.002907346396 0 jpeg:q90:djxl8 13270 4838710 2.9169834 63.213 112.574 2.17603183 0.62092407 1.811225220808 0 jpeg:q90:djxl16 13270 4838710 2.9169834 63.008 109.120 2.16807723 0.61567343 1.795909190468 0 jpeg:yuv420:q90 13270 3660915 2.2069577 93.729 238.752 6.07282543 0.95039908 2.097490592679 0 jpeg:yuv420:q90:djxl8 13270 3660915 2.2069577 93.761 156.263 6.05725956 0.91959307 2.029503031911 0 jpeg:yuv420:q90:djxl16 13270 3660915 2.2069577 95.847 152.478 6.05378914 0.91638444 2.022421724515 0 jpeg:libjxl 13270 3197015 1.9272988 18.429 99.632 2.19571280 0.70762796 1.363810502985 0 jpeg:libjxl:djxl8 13270 3197015 1.9272988 18.652 86.314 2.25490046 0.70510193 1.358942071174 0 jpeg:libjxl:djxl16 13270 3197015 1.9272988 18.639 84.488 2.29413509 0.69006634 1.329964006217 0 ``` Benchmark after: ``` Encoding kPixels Bytes BPP E MP/s D MP/s Max norm pnorm BPP*pnorm Bugs ------------------------------------------------------------------------------------------------------------------------ jpeg:q90 13270 4838710 2.9169834 64.733 166.560 2.40070152 0.68663652 2.002907346396 0 jpeg:q90:djxl8 13270 4838710 2.9169834 63.666 105.345 2.17603183 0.62092407 1.811225220808 0 jpeg:q90:djxl16 13270 4838710 2.9169834 63.061 104.653 2.16807723 0.61567343 1.795909190468 0 jpeg:yuv420:q90 13270 3660915 2.2069577 96.172 246.808 6.07282543 0.95039908 2.097490592679 0 jpeg:yuv420:q90:djxl8 13270 3660915 2.2069577 97.151 156.914 6.05725956 0.91959307 2.029503031911 0 jpeg:yuv420:q90:djxl16 13270 3660915 2.2069577 97.868 153.058 6.05378914 0.91638444 2.022421724515 0 jpeg:libjxl 13270 3197015 1.9272988 18.323 97.172 2.19571280 0.70762796 1.363810502985 0 jpeg:libjxl:djxl8 13270 3197015 1.9272988 18.660 76.811 2.25490046 0.70510193 1.358942071174 0 jpeg:libjxl:djxl16 13270 3197015 1.9272988 18.711 76.222 2.29413509 0.69006634 1.329964006217 0 ``` * Dont save MCU coding state if there are enough bytes in buffer. Benchmark before: ``` Encoding kPixels Bytes BPP E MP/s D MP/s Max norm pnorm BPP*pnorm Bugs ----------------------------------------------------------------------------------------------------------------------- jpeg:q90:djxl8 13270 4838710 2.9169834 65.817 112.747 2.17603183 0.62092407 1.811225220808 0 jpeg:yuv420:q90:djxl8 13270 3660915 2.2069577 96.322 164.534 6.05725956 0.91959307 2.029503031911 0 jpeg:libjxl:djxl8 13270 3197015 1.9272988 19.046 79.233 2.25490046 0.70510193 1.358942071174 0 ``` Benchmark after: ``` Encoding kPixels Bytes BPP E MP/s D MP/s Max norm pnorm BPP*pnorm Bugs ----------------------------------------------------------------------------------------------------------------------- jpeg:q90:djxl8 13270 4838710 2.9169834 66.330 118.045 2.17603183 0.62092407 1.811225220808 0 jpeg:yuv420:q90:djxl8 13270 3660915 2.2069577 96.814 169.467 6.05725956 0.91959307 2.029503031911 0 jpeg:libjxl:djxl8 13270 3197015 1.9272988 18.882 82.873 2.25490046 0.70510193 1.358942071174 0 ``` * Fix JPEG reconstruction error. Fixes libjxl#1790. * tools/libjxl_test.c: avoid function declaration without prototype (libjxl#1867) Avoid declaring int main() without a prototype as this is deprecated in all versions of C. * Make jpeg decoder more self-contained. (libjxl#1866) Add specialized 8x8 IDCT and transpose from jpeg xl. * Encode intrinsic dimensions (libjxl#1860) Make the encoder store the `intrinsic_xsize` and `intrinsic_ysize` from JxlBasicInfo. Setting these previously had no effect. * Add a build mode without box decoding and jpeg transcoding. In this case brotli decoder is also not needed. Brotli compressed WASM binary size is 155k. * XYB-jpeg: jpeg-specific clustering and more progression. (libjxl#1875) Benchmark before: ``` Encoding kPixels Bytes BPP E MP/s D MP/s Max norm pnorm BPP*pnorm Bugs ------------------------------------------------------------------------------------------------------------------ jpeg:libjxl:d1.0 13270 3197015 1.9272988 5.469 29.993 2.19093180 0.70477101 1.358304300765 0 ``` Benchmark after: ``` Encoding kPixels Bytes BPP E MP/s D MP/s Max norm pnorm BPP*pnorm Bugs ------------------------------------------------------------------------------------------------------------------ jpeg:libjxl:d1.0 13270 3163678 1.9072018 4.031 25.939 2.19093180 0.70477101 1.344140529099 0 ``` * fast_lossless: rename FastLosslessEncode Adding a "Jxl" prefix disambiguates it when combining it with e.g. a fast PNG encoder or fast WebP encoder. Some lines were subsequently re-formatted to fit within 80 columns. Also wrap an "extern C" around that function. Also wrap an anonymous namespace around the implementation, so that the fjxl BitWriter doesn't conflict with e.g. fpnge's BitWriter. Also fix some -Wunknown-pragmas and -Wsign-compare g++ warnings. Fixes libjxl#1855 * lint and author file * Simplex-search optimized xyb-jpeg quantization matrix parameters. (libjxl#1877) Benchmark before: ``` Encoding kPixels Bytes BPP E MP/s D MP/s Max norm pnorm BPP*pnorm Bugs ------------------------------------------------------------------------------------------------------------------ jpeg:libjxl:d0.5 13270 4804698 2.8964795 3.723 22.242 1.56636882 0.47385977 1.372525138424 0 jpeg:libjxl:d0.8 13270 3636329 2.1921362 4.089 25.696 1.87652230 0.61543190 1.349110553678 0 jpeg:libjxl:d0.9 13270 3378595 2.0367630 3.790 23.042 2.17102480 0.66006343 1.344392804990 0 jpeg:libjxl:d1.0 13270 3163678 1.9072018 4.064 24.490 2.19093180 0.70477101 1.344140529099 0 jpeg:libjxl:d1.1 13270 2984772 1.7993495 3.966 26.363 2.11484838 0.75048089 1.350377429818 0 jpeg:libjxl:d1.2 13270 2828203 1.7049630 4.278 28.153 2.29096746 0.78898283 1.345186506645 0 jpeg:libjxl:d1.5 13270 2447451 1.4754292 4.083 26.287 3.02021217 0.91575531 1.351132125102 0 jpeg:libjxl:d2.0 13270 2015903 1.2152734 4.220 31.476 3.28036547 1.09683000 1.332948351442 0 Aggregate: 13270 3064488 1.8474056 4.023 25.829 2.25605616 0.73004342 1.348686280677 0 ``` Benchmark after: ``` Encoding kPixels Bytes BPP E MP/s D MP/s Max norm pnorm BPP*pnorm Bugs ------------------------------------------------------------------------------------------------------------------ jpeg:libjxl:d0.5 13270 4654354 2.8058457 3.971 20.967 1.54231071 0.46554951 1.306260075416 0 jpeg:libjxl:d0.8 13270 3482816 2.0995920 4.085 24.342 1.70935893 0.61929160 1.300259661216 0 jpeg:libjxl:d0.9 13270 3223473 1.9432488 4.163 22.163 1.84545839 0.67493955 1.311575456149 0 jpeg:libjxl:d1.0 13270 3028977 1.8259982 4.598 25.571 1.90735793 0.71123393 1.298711878895 0 jpeg:libjxl:d1.1 13270 2833331 1.7080543 4.518 26.407 2.19600940 0.76419471 1.305286098499 0 jpeg:libjxl:d1.2 13270 2694178 1.6241669 3.971 25.394 2.27468419 0.79893476 1.297603396949 0 jpeg:libjxl:d1.5 13270 2307256 1.3909136 4.331 26.717 2.60481334 0.94243889 1.310851074545 0 jpeg:libjxl:d2.0 13270 1894603 1.1421485 4.573 33.426 3.18599176 1.12675147 1.286917548072 0 Aggregate: 13270 2919735 1.7601424 4.269 25.400 2.10445561 0.73980436 1.302161048110 0 ``` * don't compile debug print tree code (libjxl#1881) * Read the CICP tag from ICC profiles if present * Remove xyb-jpegs dependence on JXL's parameterized quant matrix. Optimize the set of remaining 105 free AC quant weights (since we want the quant matrix to be symmetric). Benchmark before: ``` Encoding kPixels Bytes BPP E MP/s D MP/s Max norm pnorm BPP*pnorm Bugs ------------------------------------------------------------------------------------------------------------------ jpeg:libjxl:d0.5 13270 4654354 2.8058457 3.663 20.463 1.54231071 0.46554951 1.306260075416 0 jpeg:libjxl:d0.8 13270 3482816 2.0995920 3.796 25.002 1.70935893 0.61929160 1.300259661216 0 jpeg:libjxl:d0.9 13270 3223473 1.9432488 3.822 22.895 1.84545839 0.67493955 1.311575456149 0 jpeg:libjxl:d1.0 13270 3028977 1.8259982 3.726 21.924 1.90735793 0.71123393 1.298711878895 0 jpeg:libjxl:d1.1 13270 2833331 1.7080543 3.716 22.769 2.19600940 0.76419471 1.305286098499 0 jpeg:libjxl:d1.2 13270 2694178 1.6241669 3.751 23.004 2.27468419 0.79893476 1.297603396949 0 jpeg:libjxl:d1.5 13270 2307256 1.3909136 4.096 27.922 2.60481334 0.94243889 1.310851074545 0 jpeg:libjxl:d2.0 13270 1894603 1.1421485 3.831 22.792 3.18599176 1.12675147 1.286917548072 0 Aggregate: 13270 2919735 1.7601424 3.798 23.258 2.10445561 0.73980436 1.302161048110 0 ``` Benchmark after: ``` Encoding kPixels Bytes BPP E MP/s D MP/s Max norm pnorm BPP*pnorm Bugs ------------------------------------------------------------------------------------------------------------------ jpeg:libjxl:d0.5 13270 4604435 2.7757524 3.217 18.406 1.67270899 0.46916401 1.302283114386 0 jpeg:libjxl:d0.8 13270 3451612 2.0807808 3.843 21.973 1.71237743 0.62384150 1.298077432889 0 jpeg:libjxl:d0.9 13270 3189961 1.9230463 3.432 19.557 1.81622076 0.67879120 1.305346906713 0 jpeg:libjxl:d1.0 13270 2999278 1.8080944 3.483 20.767 1.95959663 0.71370775 1.290450959433 0 jpeg:libjxl:d1.1 13270 2801626 1.6889412 3.668 22.541 2.30924606 0.76993853 1.300380906241 0 jpeg:libjxl:d1.2 13270 2650997 1.5981355 3.619 22.952 2.31698585 0.81157827 1.297012067011 0 jpeg:libjxl:d1.5 13270 2280336 1.3746851 3.652 24.685 2.72100186 0.94873410 1.304210588039 0 jpeg:libjxl:d2.0 13270 1875238 1.1304745 3.880 27.784 3.36580706 1.13450375 1.282527537475 0 Aggregate: 13270 2887250 1.7405590 3.593 22.166 2.17430696 0.74545951 1.297516262872 0 ``` * [butteraugli] Add support for writing the distmap to pfm. * plugins/gimp: Support for decoding animations (libjxl#1863) * plugins/gimp: Support for decoding animations * lint * fix authors check * plugins/gimp: Handle the two blend modes GIMP understands. * plugins/gimp: Handle named frames * plugins/gimp: Fix maths fail * plugins/gimp: Start layer index from 1 to match WebP/gif plugins * set coalescing to true Co-authored-by: Moritz Firsching <firsching@google.com> * ssimulacra2: linear downscaling, fixed SSIM formula (libjxl#1848) * Reapply the noise LUT fix from libjxl#1238 to the render pipeline (libjxl#1893) * Use libjpeg's source manager api in streaming jpeg decoder. This is a step towards having the same API as libjpeg. * Use jpeg_decompress_struct fields for image dimensions. * Make jpeg decoder's API more similar to libjpeg's API. Modify ReadHeaders() to read until SOS marker as in libjpeg. Add a StartDecompress() API function that will read whole input for progressive images. This is the same behaviour as in libjpeg's jpeg_start_decompress() for non-buffered-image mode. * plugins/gimp: Fix progressive decoding (closes libjxl#1845 and libjxl#638) (libjxl#1870) * Refactor WASM demo code * Update dec_xyb.cc to fix wasm decode issue (libjxl#1900) * Update dec_xyb.cc Changed OR to AND. In WASM, images that were not XYB encoded would fail to decode, as `CanOutputToColorEncoding(c_desired)` is never called. * fix authors check Co-authored-by: Moritz Firsching <firsching@google.com> * Fix the automatic setting of intensity_target from the input colorspace (libjxl#1899) * Use Set(Quality|Distance) where appropriate Goal: fix gcc-10 warning SetQuality is unused. * Add demo site generator * Add option to run custom cjpeg command in jpeg benchmark. (libjxl#1905) With this addition, one can benchmark against a decoder-only libpeg implementation by setting a custom LD_LIBRARY_PATH and using a statically built cjpeg in the benchmark. * Fix WASM test * First version of decoder-only jpegli library. (libjxl#1909) The jpegli library is a drop-in replacement to libjpeg62, with improvements inspired by the JPEG XL codec. The commit adds a minimal version of the libjpeg decoding API. After building the jpeg and benchmark_xl targets, the following command can be used to convert an input.jpg file to PNG: ``` LD_PRELOAD=./build/libjpeg.so.62 ./build/tools/benchmark_xl --input input.jpg --codec=jpeg --decode_only --save_decompressed --output_dir . ``` * Support 16 bit output buffer in libjpegli decoder. (libjxl#1913) This can be enabled by the application by setting ``` cinfo.quantize_colors = FALSE; cinfo.desired_number_of_colors = 65536; ``` In this case a scanline buffer should be twice as long. Other changes in the commit: * Remove code duplication between input reading API functions. * Use the existing cinfo fields for marker parsing and keeping internal state, remove redundant fields from jpeg_decomp_master. * Add trivial implementation of jpeg_calc_output_dimensions and use cinfo->output_{width,height}, cinfo->out_color_components where appropiate. * Update HWY Add WASM-256 to build matrix * Refactor ServiceWorker code ... to make it readable / hackable. * libjpegli: reuse more fields of cinfo, fix build. * mention Pale Moon support (libjxl#1915) closes libjxl#1914 * Disable libjpegli on release and debian package builds. * Avoid ODR violations * Ensure alignment in OpHlg * mention GraphicsMagick support (libjxl#1923) * jpegli: fix decoding of non-interleaved sequential jpegs. (libjxl#1922) * Add 16 bit jpeg output mode to benchmark_xl and cjxl. (libjxl#1918) If the benchmark_xl and cjxl binaries are run with libjpegli, then they will decode jpegs to 16 bit ppfs, otherwise they work as before. * Read cICP chunks from PNGs - Where possible, create an encoded profile from the cICP values instead of using an ICC profile. - Change default rendering intent for decoded PNGs to "Relative". - Honor the rendering intent specified by sRGB chunks when applicable. * Explicitly set sRGB profile when reading the chunk To ensure that sRGB takes priority when cHRM/gAMA occur earlier in the file. * fix hshift/vshift of metachannels (libjxl#1925) * cjpeg_hdr: reuse EncodeJpeg Co-authored-by: Luca Versari <veluca@google.com> * Add a basic quality setting to cjpeg_hdr Co-authored-by: Luca Versari <veluca@google.com> * Fix NEON SIMD implementation of fast_lossless Also enable SIMD in build scripts. * remove deprecated or dead code * jpegli: implement jpeg_consume_input() Added tests for suspending source manager use-case, and tests for consuming input ahead of the output processing. * Refactor fast_lossless to support 16-bit data. >8 bit paths are not SIMDfied yet. * premultiply alpha in BufferToImageBundle (libjxl#1927) * premultiply alpha in BufferToImageBundle * remove premultiply_alpha arg * remove arg also from SetAlpha * don't include *_gbench.cc in code coverage report (libjxl#1939) * Fix asan build. * don't generate unused noise (libjxl#1943) Fixes libjxl#1888 * Fix ExternalImageTest asan failure. * Add an --update_on_failure flag to the conformance test runner. Also update the conformance test set commit to make the tests pass. * jpegli: implement buffered image mode * jpegli: implement decoding to cropped output. The implementation is not yet optimal, it just calls jpeg_read_scanlines and discards the output that is not needed. * Build release package for Ubuntu 22.04 * jpegli: implement raw data output mode Drive-by: fix bug in bias computation for buffered image mode and 32-bit architecture. * Add libjpegli's libjpeg.so target to build/lib/jpegli directory. * SIMDfy >8-bit fjxl on ARM NEON. Benchmark on a A51, approximately a 10% speedup on a single core: ``` Before: ./fast_lossless artificial8.ppm x.jxl 2 100 63.178 MP/s 2.755 bits/pixel ./fast_lossless artificial13.ppm x.jxl 2 100 32.670 MP/s 6.994 bits/pixel ./fast_lossless artificial14.ppm x.jxl 2 100 32.285 MP/s 7.956 bits/pixel ./fast_lossless artificial16.ppm x.jxl 2 100 27.592 MP/s 10.411 bits/pixel ./fast_lossless bb8.ppm x.jxl 2 20 37.628 MP/s 11.895 bits/pixel ./fast_lossless bb13.ppm x.jxl 2 20 21.407 MP/s 26.594 bits/pixel ./fast_lossless bb14.ppm x.jxl 2 20 20.894 MP/s 29.654 bits/pixel ./fast_lossless bb16.ppm x.jxl 2 20 18.621 MP/s 35.989 bits/pixel After: ./fast_lossless artificial8.ppm x.jxl 2 100 63.353 MP/s 2.755 bits/pixel ./fast_lossless artificial13.ppm x.jxl 2 100 36.205 MP/s 6.994 bits/pixel ./fast_lossless artificial14.ppm x.jxl 2 100 34.875 MP/s 7.956 bits/pixel ./fast_lossless artificial16.ppm x.jxl 2 100 29.342 MP/s 10.411 bits/pixel ./fast_lossless bb8.ppm x.jxl 2 20 38.161 MP/s 11.895 bits/pixel ./fast_lossless bb13.ppm x.jxl 2 20 23.457 MP/s 26.594 bits/pixel ./fast_lossless bb14.ppm x.jxl 2 20 22.861 MP/s 29.654 bits/pixel ./fast_lossless bb16.ppm x.jxl 2 20 20.144 MP/s 35.989 bits/pixel ``` Co-authored-by: Zoltan Szabadka <szabadka@google.com> Co-authored-by: Leo Izen <leo.izen@gmail.com> Co-authored-by: szabadka <9074039+szabadka@users.noreply.github.com> Co-authored-by: alistair7 <alistair7@users.noreply.github.com> Co-authored-by: Nigel Tao <nigeltao@golang.org> Co-authored-by: Moritz Firsching <firsching@google.com> Co-authored-by: Jon Sneyers <jon@cloudinary.com> Co-authored-by: Sami Boukortt <sboukortt@google.com> Co-authored-by: Luca Versari <veluca@google.com> Co-authored-by: Fredrick Brennan <copypaste@kittens.ph> Co-authored-by: Evgenii Kliuchnikov <eustas@google.com> Co-authored-by: Jim Robinson <jimbo2150@gmail.com> Co-authored-by: Damiano Albani <damiano.albani@gmail.com>
jonsneyers
pushed a commit
to jonsneyers/libjxl
that referenced
this pull request
Apr 24, 2023
Benchmark before: ``` Encoding kPixels Bytes BPP E MP/s D MP/s Max norm pnorm BPP*pnorm Bugs ------------------------------------------------------------------------------------------------------------------ jpeg:libjxl:d1.0 13270 3197015 1.9272988 5.469 29.993 2.19093180 0.70477101 1.358304300765 0 ``` Benchmark after: ``` Encoding kPixels Bytes BPP E MP/s D MP/s Max norm pnorm BPP*pnorm Bugs ------------------------------------------------------------------------------------------------------------------ jpeg:libjxl:d1.0 13270 3163678 1.9072018 4.031 25.939 2.19093180 0.70477101 1.344140529099 0 ```
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Benchmark before:
Benchmark after: