lossless mode is not bit-exact and the result is different on every run

[Elyotna]

Setup :

x265 [info]: HEVC encoder version 1.7+418-d56b2466c044

x265 [info]: build info [Linux][GCC 4.9.2][64 bit] 8bit

2x Intel® Xeon® E5 2660 v3 - total 20c / 40t

Bug :

The --lossless option produces non-deterministic outputs that are (sometimes) not bit-exact lossless : the MD5 sums of the raw input and the raw decoded output don't match.

How to reproduce :

Encode any sample with --lossless. Tested presets were ultrafast, superfast and veryfast.

Additional notes :

The non-deterministic outputs seem to be more chaotic the more threads are added through --numa-pools.

Sometimes, the MD5 sum will match and the output will be indeed lossless. If that is the case, run the test once again until you get another output that probably won't match the MD5 sum.

[Elyotna]

A sample that will reproduce the issue : https://media.xiph.org/video/derf/y4m/ducks_take_off_420_720p50.y4m

bug reproduced on my setup using numa-pools=20 and preset ultrafast

More info : the bug does not seem to be in 1.7. If you are interested, I could try to bisect it out.

[Deepthi Nandakumar]

Hey, thanks for reporting - we will look into it. I'm guessing the qg-size fix wasnt enough.

[Peter Steinbach]

as this is an important feature for my use case, I tried to reproduced the error given the above dataset. with no numa pools used, I coudn't reproduce it with 1.7+425-2980141a744a.

[Deepthi Nandakumar]

Our tests show the error has been fixed with 575bf6d52e15. Please re-open if this is not the case.