Control: found -1 1.2~alpha2-1 Control: tags -1 + pending Hi,
On Tue, May 15, 2018 at 09:24:05AM +0200, Steinar H. Gunderson wrote: > PLC is all broken for CELT in 1.2.1 due a patch that optimized celt_fir() for > ARM; > it mostly outputs just silent frames. (You can reproduce it by giving -loss to > opus_demo.) It looks like the patch which caused this regression was two commits before the 1.2-alpha2 tag, which we shipped for Stretch, so I'm marking this as found there too, in case someone else hits it there later. > Until 1.3 comes out, which could be a while, would you be willing > to carry this patch from upstream? > > > https://gitlab.xiph.org/xiph/opus/blob/072d133f7899c4783e67f90d07ab25b3b8414b8f/celt/celt_lpc.c > > There's one more patch that looks relevant > (92ffce621df6ace95267ac8c13aa0d862f6a476b), > but I haven't tried it, and at least for my limited testing, the first one > seems > sufficient. I just missed catching you in #opus, which might have saved you repeating what you found here, if I'd seen that before I saw this :) But at least we have a report for anyone who might also hit it in Stretch, and this is what we decided there after you'd gone: The upstream plan is to tag at least a 1.3-rc "sometime soon", and at present there are no known regressions in what's currently committed to that branch - so since we've got some breathing space before the next freeze, then rather than cherry-picking just that one patch, I'll push out a snapshot from git head which contains it and all the other fixes to date, including the new 'hardening' sanity checks and the tuning improvements. Then we can get some extra testing on those too, which will increase the confidence for making a formal 1.3 release if nothing extra shakes out of it. Thanks for digging into this and reporting it. Cheers, Ron
