Files are here:
https://sourceforge.net/projects/libjpeg-turbo/files/1.5.90%20%282.0%20beta1%29/
Change log is here:
https://github.com/libjpeg-turbo/libjpeg-turbo/releases/tag/1.5.90
YUM repository is unfortunately not updated yet and won't be until
SourceForge restores SSH access:
https://sourceforge.net/p/forge/site-support/16424
(NOTE: If anyone knows of another file release system that supports
interactive SSH and rsync access and provides download stats, please let
me know. The only thing we're still using on SourceForge is the file
release system, but I haven't yet found another one that meets the needs
of this project. Refer to
https://github.com/VirtualGL/virtualgl/issues/35 for discussion.)
This release was made possible by a grant from the Mozilla Open Source
Support (MOSS) program.
Release Highlights:
-- AVX2 (256-bit) SIMD acceleration for the colorspace conversion,
chroma down/upsampling, integer quantization/sample conversion, and
"slow" integer DCT/IDCT algorithms. This speeds up the overall
performance of the library by about 15-20% on a Core i7 ("Haswell") CPU
when using the "slow" integer DCT/IDCT (the "slow" integer DCT/IDCT now
performs similarly to the "fast" integer DCT/IDCT on that platform.)
Looking for feedback regarding the performance on other AVX2-equipped
CPUs. Refer to
https://github.com/libjpeg-turbo/libjpeg-turbo/issues/2
for discussion.
-- SSE2 SIMD acceleration for progressive Huffman encoding. This speeds
up the overall performance of progressive JPEG compression by almost 2x
on SSE2-equipped CPUs.
-- Complete build system overhaul. We are now using CMake and only
CMake for all platforms. Refer to
https://github.com/libjpeg-turbo/libjpeg-turbo/issues/56 for discussion.
-- ICC profile read/write support in the libjpeg API
-- Improved error handling in the TurboJPEG API (thread-safe error
retrieval, optionally making libjpeg warnings fatal, determining whether
a function failed due to a libjpeg warning or a fatal error.)
-- BMP/PPM load/save support in the TurboJPEG API
-- Loongson MMI SIMD acceleration for select algorithms, which speeds up
compression by ~70-100% and decompression by ~2-3.5x.