Reported by DRC: A user reported, and I reproduced, a severe accuracy problem in the libjpeg-turbo fast integer forward DCT SIMD code that occurs only when the JPEG quality is >= 98. For these JPEG qualities, the SIMD version of the code ceases to produce identical results to the non-SIMD version of the code, which I consider to be a bug. More info: Further information on this: I've narrowed down the problem to the jsimd_quantize_sse2() and jsimd_quantize_mmx() functions in jcqnts2i[-64].asm and jcqntmmx.asm, although I can't see any obvious mathematical errors in that code. And SF tracker entry: http://sourceforge.net/tracker/?func=detail&aid=3165031&group_id=303195&atid=1278158
This bug has been fixed upstream, and will be solved if we upgrade libjpeg-turbo.
(In reply to comment #1) > This bug has been fixed upstream, and will be solved if we upgrade > libjpeg-turbo. Correction: This was fixed in 1.1.0 (ref: https://github.com/libjpeg-turbo/libjpeg-turbo/releases/tag/1.1.0) and we are using 1.5.3.
Nothing specific to test here, and we already did general testing of libjpeg during the upgrades.
